Cracking the code: How generative biology could transform medicine

Can AI unravel the mysteries of human biology? Could it help design specialty treatments and cures for disease? Geoffrey von Maltzahn and his team at Generate:Biomedicines are bullish on both counts. AI has greatly accelerated progress in genome engineering, bioengineering and nanotechnology and they are getting closer to developing tailored therapeutics. “Six years ago, this was a crazy idea,” he says. “We’re now convinced that 100 percent of protein therapeutics are going to get created this way.” In this special episode of Solve for X, host Manula Selvarajah sits down with von Maltzahn to talk about where the science is now and where it is headed.

Subscribe to Solve for X

 
Geoffrey von Maltzahn: Machines don’t have many of our limitations. And about six years ago, we started to hypothesize that the future of biology would have a whole different force behind its creation.

Narration: When it comes to human biology, it’s almost as if Geoffrey von Maltzahn has a front-row seat to the future. He’s someone who can see how machine learning and AI will radically transform life sciences.

Geoffrey von Maltzahn: And that was generative AI, mastering patterns that are resident in the codes of life, the structures of life, figuring out how to create the future.

Narration: Geoffrey works at the vanguard of biology. At Flagship Pioneering — he’s launching revolutionary biotech startups. You may have heard of one of their biggest: Moderna. He has founded more than 10 companies, including Generate:Biomedicines, a startup that uses AI to discover new drugs. Machines, as Geoffrey explains, can work at speeds far beyond what humans are capable of. They can help us untangle some of life’s great mysteries to understand how the living world works and figure out things that have eluded scientists for centuries.

I’m Lara Torvi, a producer at Solve for X: Innovations to Change the World. Geoffrey von Maltzahn gave a keynote at the MaRS Impact Health conference about the mechanics, challenges and limitless potential of generative biology. Afterwards, our host Manjula Selvarajah had the chance to chat with him.

Manjula Selvarajah: Hi Geoff.

Geoffrey von Maltzahn: Hey there.

Manjula Selvarajah: Most of your work focuses on the convergence of biology and artificial intelligence, or what’s called “generative biology.” How would you explain what generative biology is to a fifth grader?

Geoffrey von Maltzahn: I have a one-year-old, four-year-old, six-year-old, eight-year-old at home. So if I, if I miss a fifth grader and land on a first grader…

Manjula Selvarajah: We’ll take it. We’ll take it.

Geoffrey von Maltzahn: Forgive me, eventually I’ll have to do that. So first, let’s talk about, what does the word generate mean? And let me start in the realm of images. I think most people are familiar with artificial intelligence recognizing patterns to do facial recognition. So I take a picture of you, and then I’d be able to recognize you in various streams of video instantly based on patterns that the machine uniquely associates with your face. To generate, isn’t recognizing something that exists, but it means to create something new. And so similar models, but with a distinct intent and construct, can learn patterns that are universal to human faces and start to generate new human faces of people who’ve never existed with sufficient detail now that statistically, those people will never exist, and they never have, but would be unmistakable to you or I from a picture of a person walking around in the world today. So that is to generate — it’s to learn about what the fundamental patterns of a given object, entity or language are, so as to be able to utilize those patterns and create new things. When you bring that into the world of biology, what that means is machines learning about a category of biology with sufficient fluency, they can generate new examples when challenged to achieve some desired technological result.

Manjula Selvarajah: Where are some of these places, or what are some of these places where we’re seeing this in practice?

Geoffrey von Maltzahn: So we had this idea that you might be able to generate biology. We had to choose where one would be able to start in biology, and we fell in love with proteins for a couple of reasons. One, is proteins have a luxury of two very large and high-quality data sets. Which, in biology, large high-quality data sets are a scarce thing. One, is that all proteins come from DNA. So, the human genome encodes 20 plus thousand proteins. A bacterial genome encodes thousands of proteins. And for 20 plus years, the life sciences technology community has been sequencing genomes across every branch of the tree of life. And that primary code was a wonderful way to start to let machines try to learn what are the relationships between DNA code and various categories of proteins with function, with the hope that they might become Shakespearean in learning how to write an antibodies code or an enzymes code, et cetera. The second large data set was all crystallized 3D structures of proteins. So almost every other week, in the preeminent journals in science, a new three-dimensional protein structure has been present for the last 30 plus years. And those have been the result of a huge amount of painstaking effort to try to purify, isolate and crystallize that protein so that we can figure out what three-dimensional shape exists in nature. That’s a community that has been obsessing over the quality of that data, and there’s been a big incentive to figure out, “What are the motors and sensors and machines in the protein realm of life look like at a molecular scale?” So, we wondered whether applying generative AI to learn the rules of DNA and to learn some of the ways that amino acids like to interact with each other might allow us to do new things. Among the most interesting things that we’ve been able to do is in that second category. It may sound simple, but figuring out how amino acids like to interact with each other is effectively the most valuable thing in all of biotechnology. All antibodies today, all peptides today, all cytokines, growth factors, they have one function, which is upon entering a patient, to go and stick to a target at a really specific location and not stick to other stuff. So if you could figure out how amino acids like to interact with each other, you might be able to fundamentally change the way that every single one of those gets created in the future. Most of those have been sort of fishing expeditions into really inefficient discovery processes, including what our biology does when it needs to make an antibody.

Manjula Selvarajah: So that’s interesting. I love that terminology that you use. It is fishing, and it is a long expedition, right? With a lot of failure rates, lengthy time that this happens. Which part of it do you see changing the fastest? Is it the discovery stage? Talk to me about which part you think will change with generative biology.

Geoffrey von Maltzahn: I’m closest to the discovery side, and it’s going to change really quickly. On a relative basis, other things are going to change also, but I don’t have as clear a picture as to over what time scale. But six years ago, this was a crazy idea. Four years ago, we became convinced that this was going to work. We’re now convinced that 100 percent of protein therapeutics are going to get created this way: In that, that fishing expedition of looking for things with trial and error in nature or in the lab, as you say, is a really, really long expedition. Let me kind of put some numbers around it to put into perspective that it’s effectively an infinite expedition. Let’s take an antibody, for example. Antibodies, when they stick to a target, they use about 60 (six zero) amino acids to do so. Every amino acid, every one of those positions, has 20 options. So the number of potential antibodies at the domain that matters for what it’s binding is 20 to the 60 power. That’s more than all the atoms in the entire universe. So, if you had every atom on Earth, you could only build an infinitesimally small fraction of what’s possible. So it’s more miraculous that you and I, and any one of us, when we get sick, ever find an antibody that works, than it is to imagine that there isn’t any antibody in nature, which is really optimized. Or that, you know, at least some of the things that we rely upon in medicine today have limitations that this kind of technology is going to be able to overcome.

Manjula Selvarajah: It’s interesting to hear that that your thinking has moved from a time where you thought, “I don’t know if it’s possible”, or you and your team, “I don’t know if this is possible” in six years to, “What are we going to target next?” In a way, right? I think that’s really fascinating. And that gets me… Actually I wonder, is this then a turning point? Do you see us in the next while experiencing this sort of leap forward in drug discovery and treatment discovery?

Geoffrey von Maltzahn: Yeah, and I’ll give you two examples. Fundamentally, the reason I was inspired to go into biology is, by contrast, if you’re an engineer and you build a bridge, and one of your bridges falls down? Like, you’re in trouble. In biology, if you build a million biological bridges and one of them stands up, you might win a Nobel Prize. The sort of victories in biology are so scarce and so hard fought, which just, you know, gives a sense of how high the ceiling may be of what we can do. And it looks like these tools are going to allow us to do things that we can already do, but much better, and do things that we can’t do at all today. So those are represented by the first two antibodies that we brought into the clinic. The first is, a couple of years ago, this is now three years ago, we challenged the computational team at Generate to take the top $50 billion of antibody therapeutic sales and Generate antibodies would hit the exact same target, same epitope, which means spot on the target, same…

Manjula Selvarajah: And these are commercially available ones…

Geoffrey von Maltzahn: Yeah. These are the most valuable antibodies in all of medicine.

Manjula Selvarajah: OK.

Geoffrey von Maltzahn: Same structure at the interface, comparable or better affinity, which means how strongly they stick to the target, without going anywhere near the intellectual property of those parents.

Manjula Selvarajah: Of course, of course.

Geoffrey von Maltzahn: So in three months, the team was able to do it for 100 percent of them, and you wouldn’t be able to do that in a decade, if you had all pharma companies teams together.

Manjula Selvarajah: That’s pretty impressive.

Geoffrey von Maltzahn: So what that means is that when an antibody shows that it can help patients, medicine may potentially be able to access the very best antibody for that spot. Sometimes the first one is the best, but usually it has limitations: either dosing frequency or its potency for the target or our immune system responds negatively against that antibody. And things that can be improved upon to make a much better and more valuable medicine. And, you know, those criteria I just described is almost like Robin Hood splitting a first arrow in the bull’s eye 50 times in a row — it just changes the ease with which we can do things. If you had been looking for similar things with conventional tools, you’d find stuff that sticks, like things on the target, but not the criteria that I’ve described. So it’s, of course, also valuable to do things that you can’t do at all, and everybody’s favorite acronym — COVID — is one of the things we’ve been paying attention to. It’d be nice if we could forget about this acronym for rational reasons.

Manjula Selvarajah: We can’t.

Geoffrey von Maltzahn: Yeah, the reason we can’t is the cuddly spike protein that sticks out of COVID has been changing voraciously on the parts that our immune system creates antibodies against. So your prior COVID infection, by the time you face the latest variant, the antibodies may not give you very much help in preventing a new infection. There are other parts of the spike protein that haven’t changed almost at all from COVID in 2019 to the wildest variants today. In fact, you can march up the tree of life of SARS viruses and these parts are still constant. The problem is your and my immune systems (and humanity’s immune systems) are terrible at finding antibodies that can hit those sections. So we asked whether generative AI could help us there. And in fact, we’ve been able to identify, generate antibodies that can hit those highly-concerted parts.

Manjula Selvarajah: The ones that stay. Oh, interesting.

Geoffrey von Maltzahn: So this has the potential to be another tool, should there be a new pandemic, or a means by which we can perhaps better abate the evolution of COVID against the tools that we’ve had previously and the ones that we have right now.

Manjula Selvarajah: So what I sense about, I mean, I felt that before our conversation, I’m certainly feeling it now, is this idea that we are at a turning point, that there is, if not already, this possible leap forward: Finding drugs, finding treatments faster. Is there anything that you think will hold that leap back?

Geoffrey von Maltzahn: So, I’m deeply optimistic by nature. I think…

Manjula Selvarajah: Pessimistic for a minute, if you could.

Geoffrey von Maltzahn: OK, I’ll give you a few counterpoints. In fact, you can’t really do things entrepreneurially unless…

Manjula Selvarajah: Yes, no kidding. That’s true. Stare into the valley and feel hope. OK.

Geoffrey von Maltzahn: There’s a great Charlie Munger quote, which is something like, “I want to know where I’m going to die so that I never go there.” That’s true for startups.

Manjula Selvarajah: (Laughing) Yes. We’re going to keep that quote.

Geoffrey von Maltzahn: We’re going to count all the ways you can die and, you know, find ways to avoid everyone.

So first let me kind of depict why this has the potential to be so marvellous. In essence, humanity’s ability to innovate has been a function of our knowledge and our intelligence. And the intelligence revolution is going to change the way that we can utilize the knowledge that we already have in extraordinary ways. And life science, I think, is like the most interesting playground for that, because of how little we’ve understood and because of how much data that we don’t understand already exists.

So there’s this backlog of, you know, almost like an Internet worth of language that, you know, we’ve learned almost nothing about relative to the majestic things that it contains. I mean, our genome contains us. So when you put our current understanding in light of that benchmark, it’s got a long way to go. As intelligence increases, it almost is like a time machine, because otherwise, things that sit 100 years in the future or decades in the future with current intelligence can be pulled into the present, and it looks like the search efficiency for finding new medicines is going to go way up, and that’s just one fraction of our healthcare system, and what it means to be healthy. AI is going to affect other ones, too.

So the things that could deter that are a combination of, once we’ve squeezed the juice of what can be done with prior data, thereafter, we’re going to be limited by how quickly we can run interesting experiments. And science isn’t very efficient in doing that. Like, you know, I’ve kind of been in the coal mine of running experiments for 20 years. And for anybody that’s been in a life science lab with a lot of time, there’s a lot of really minute and detailed and error-prone aspects of the way that the frontiers of science advance knowledge. So once all current knowledge has had models trained on it, that’s going to lead to a huge number of hypotheses that are equally plausible about what a technology could look like, or how the world really, really works. We’re going to have to be able to test those.

So I think we’re going to see a huge amount of innovation to speed up science. And without that, we’ll have this, kind of, like, burst of progress, but then maybe a dripping faucet of incremental advances that are experimentally limited, not intelligence limited. There’s also going to be a need for regulatory regimes to change, adapt. And the fact that discovery has been so slow has not put undue scrutiny on the fact that development is really slow. All of a sudden, development is going to look like a DMV-like experience relative to normal life. DMV is the Department of Motor Vehicles in the United States.

Manjula Selvarajah: Even we know what that term is. Though we don’t, I don’t think we use the same term, but yeah, so it is this idea that you have a speed boat on one side, or, or it used to be a slow boat, but everything was slow, so it didn’t matter. But now what you have is something that’s speeding ahead, which is now trying to function in a system that is used to going much slower and could go much slower.

Geoffrey von Maltzahn: I also think there’s going to be a kind of tectonic shift in the way that startups exist, in that, one of the reasons that the frontiers of science have been the home for academic labs and startups is that, historically small teams of humans efficiently communicate and can kind of bravely figure out, you know, how to move a point to the court. Now, there are scale benefits to intelligence, and the biggest companies in the 20th century were built on economies of scale. It’s probably going to be the case that the biggest companies this century will be built on a different scale effect: Intelligence of scale, where you get wiser with every single data point, whether that be chemistry data, biological data, physical science data or others. That may mean that some breed of new startups are going to become more intelligent as they scale, and you wouldn’t want to be off of that platform, which may totally change, you know, the home for the cutting edge of innovation.

Manjula Selvarajah: Let’s talk about Flagship for a minute, your business model, which is unconventional actually for a venture capital firm. So what you do is you actually take something from conceiving it through to creating it, to resourcing it, developing it. And what you develop are these things that you call bio platform companies. What is a bio platform company?

Geoffrey von Maltzahn: Yeah, so the word platform is used frequently, and sometimes its definition drifts, or it’s implied that people know. We think about a platform as a set of capabilities that are technological in nature, that allow you to be able to do a given endeavour repeatedly. So I’ll give an example of Moderna. Flagship’s most famous for having started Moderna. Moderna brilliantly stated in the early days that there would never be a Moderna that would end up with a drug. It would either be zero because it doesn’t work at all, or the number is going to be enormous in that mRNA, sitting in the center of the language of life, was unlikely to work in some tiny hallway of medicine and not elsewhere. It either would structurally be non-viable as a therapeutic or it would have applications all over the place. Moderna built a platform to solve a lot of really hard problems associated with reliably designing, producing small quantities, testing, delivering and then eventually manufacturing a whole new category of therapeutics. It is easier for them to take on any other mRNA endeavour than it is for any other entity on the planet. And that’s the way we think about what a great bio platform looks like: If you invent a set of capabilities that then allow you to create a new kind of product.

Manjula Selvarajah: It’s interesting, because when I think of these kind of companies, I’m trying to imagine what it takes to, like, what kind of talent and infrastructure you have to have at Flagship in order to be able to… there’s 10 companies now, 10 of these bio platform companies under Flagship. Is that about right?

Geoffrey von Maltzahn: It’s closer to 100 in the history of Flagship.

Manjula Selvarajah: OK, so what does it take…

Geoffrey von Maltzahn: That probably exacerbates your question.

Manjula Selvarajah: Yeah. What does that, I mean, what does that look like then, in terms of the talent and infrastructure you need to have Flagship run and do all of these things?

Geoffrey von Maltzahn: There’s a bunch of components to it, and we’re still figuring it out. So, a big part of it is a belief that extraordinary entrepreneurs can be trained. Sometimes there’s just like nature versus nurture. When, in fact, what entrepreneurial people are trying to accrue is like 12 professions of competence and mastery at the same time. So you could get 12 graduate degrees, you know — three legal ones, two technical ones, an MD and more — you still wouldn’t have figured out how to weave it together. And so, oftentimes entrepreneurial people are, like, left with almost destructive advice, like, “Oh, you’re really entrepreneurial. The best thing you can do is just go off on your own.” And if you’re just trying to do one of those professions, it would be bad advice to learn it online. Somehow doing 20 of them, that becomes the default, and it’s in part because there aren’t places where you can really train those skills.

So from the outside, Flagship’s regarded as a place that companies come from. Inside, we think of it as a place where we’re trying to train the world’s very, very best life science entrepreneurs. All of us are aspiring to become the best at starting with nothing and inventing a technology that corresponds to a bio platform that might just change the world. And we’ll lead the companies across every one of those phases, from absolutely nothing to no people, to a few people, to 100 people. We’ll then typically hand off that responsibility to somebody who is the best for the stage of the company growing up, you know: Putting therapeutics into the clinic, commercializing products in agriculture or other things. So that’s one stripe of it.

There’s a great Edwin Land quote. He was the founder of Polaroid that “If you want to be extraordinary at one thing in life, you should accept being ordinary at everything else.” And, in a sense, that affects our model, or that thinking does, in that we’re trying to allow extraordinary life science entrepreneurs to work with the world’s most thoughtful and creative intellectual property leaders and the same for legal and HR and talent, and thereby sort of endowing a wild idea, a crazy technology team with some extraordinary strengths that reduce the number of unforced errors, but also allow that composite of, you know, what ultimately is necessary for something to be a great company to be forged with the product of many people’s contributions.

So it’s a beautiful team sport. Companies are complicated life forms.

Manjula Selvarajah: So, before I let you go, you’ve talked about these hundred companies that are part of the history of Flagship, and there’s a ton of things that you’re working on in terms of biomedicines. What are some of the areas that you’re excited about?

Geoffrey von Maltzahn: I’ll give you one of the threads that I’ve been hinting at, which is, it turns out that generative AI is going to have all kinds of applications in the future of technology. So, protein therapeutics is one really big category. Small molecules are another. RNA and DNA medicines are another. But it’s also not limited to life science. We’ve started spending time thinking about ways that similar principles might learn about the physical world in ways that our human brains haven’t been able to understand. So what is it about the three-dimensional arrangement of atoms that lead it to be a great electrode? Or a great material for building applications and construction? Or something that could capture CO2 from the atmosphere or from industrial sources?

If we could learn those relationships, then we may vastly accelerate humanity’s ability to get out of some of these challenges that we’ve been creating for ourselves: the climate crisis, many forms of sustainability that are beyond our technological reach right now as well as categorically improving our health system.

Manjula Selvarajah: So take it from biology and health to other realms.

Geoffrey von Maltzahn: Yeah.

Manjula Selvarajah: Well that’s interesting. We’ll have you back to talk about it.

Geoffrey von Maltzahn: Cool, I’m looking forward to it.

Manjula Selvarajah: Geoff, thank you so much.

Geoffrey von Maltzahn: Thank you, I really appreciated the time.

Narration: Solve for X is brought to you by MaRS. Lara Torvi and Sana Maqbool are the associate producers. Mack Swain composed the theme song and all the music in this episode. Gab Harpelle is our mix engineer; Kathryn Hayward is our executive producer. I’m your host, Manjula Selvarajah. Watch your feed for the next episode next month

Solve for X is brought to you by MaRS, North America’s largest urban innovation hub and a registered charity. MaRS supports startups and accelerates the adoption of high-impact solutions to some of the world’s biggest challenges. For more information, visit marsdd.com. And we want to hear from you — drop us a line to share your ideas, questions and feedback. What innovations are you curious about? Email us at media@marsdd.com.

Illustration by Workhouse and Kelvin Li