If DNA is Code, Can AI Help Write It? Scaling Cell Programming and Synthetic Biology, with Ginkgo Bioworks Co-founder and CEO Jason Kelly

0:05Biology is undergoing a digital revolution as we build developer tools and production infrastructure for synthetic biology. How will it change industries? How does it intersect with AI? And how do we rethink biosecurity? This week, Sarah and I are joined by Jason Kelly, co-founder and CEO of Ginkgo Bioworks, to discuss their goal of making cells as easy to work with as computers, their data strategy, and the tech keeping the next pandemic at bay. And in general, what cell programming will do for the future of food, medicine, and agriculture. Jason, thanks so much for joining us today. Yeah, thanks for having me on.

0:36So I think there's a lot of talk about synthetic biology and how biology and DNA and proteins are effectively just code and you can manipulate them in different ways now and things like that. I'd love to just get your view of both what Ginkgo does as well as what does synthetic biology actually mean? So I think the founding idea of synthetic biology is that DNA is code, right? And inside of cells are A, T, Cs, and Gs, essentially on like a tape. and it is very like surprisingly analogous to zeros and ones and you know inside memory in a computer that's roughly where the similarities end okay like once you get to the next step of what the cell does with that code we are in a totally different world it is not virtual is the first thing right it is a physical thing the code itself is literally physical right it is a polymer and it is going to use that to make proteins which are basically little pieces of nanotechnology and and they're all going to be bumping into each other, and it's all crazy.

1:30It's not physically isolated like you would imagine with a semiconductor chip. It's not built by humans. So you have this really interesting thing where the hook is there for people in tech to engage with biology, but then once they get in, they're like, what the fuck? And so I'm happy to talk about those pieces, but I think you're right. The core idea of Symbio is that it runs on code, and then what can we bring over from programming into this world that actually sticks? And so I think what SynthiBiology has been, you know, really since it got going, I met, you know, the founders of Ginkgo back when we met at MIT in 2002.

2:04That was like early days of SynthiBio. It's about 20 years now. It's basically engineers asking the question of what can they bring over into biology that's actually going to work. And some stuff has been left by the wayside and some things do work. And the latest technology that's being tried now is AI. Can you walk us through what you actually think does transfer over and then where there are one or two unique challenges and then how does AI help to solve for some of those things? I'll tell you like a funny story, right? So one of the fellows I started the company with is this guy, Tom Knight, right?

2:30And Tom Knight started on the faculty at MIT in 1972. Okay, right? Like mainframe computers, punch card computers. He was a computer architect for a very famous mini computer, which was like the size of a refrigerator called the Lisp Machine. Okay, like Symbolics, that company is one of the founders of. Like old school classic Stephen Levy hackers in the book kind of guy, right? Mid-90s, he realizes this thing about DNA is code. and basically is like, forget computers, I'm moving into programming DNA. He's still Tom, right? He's been like teaching the semiconductor course for 20 years at MIT at this point.

3:03Opens a wet lab in the MIT computer science building. Starts growing bacteria, freaking everybody out, right? And he puts up this flag and he's like, hey, computer scientists, like DNA is code. If you're interested in this thing, like come over and try it out, right? Some of us came over and we're like, all right, cool. We got there, got our hands wet and we're okay with it. A lot of computer scientists, they get there. Tom's like, okay, here's the lab bench. Remember, this code is physical. So if you want to compile it, I'm going to have to teach you how to do molecular cloning. And here is a pipette.

3:31And you're going to sit at this bench and you're going to do these steps, okay? And the person would do them. And they'd get a result the next day. They're like, wow, that's really interesting. And then they'd do the same thing again the next day and they would get a different result. And they'd be like, Tom, I just did the exact same thing and I got two different results. Like, what's going on? And he was like, you're never going to know. and it would just break their brains, right? Like they'd be like, fuck this, I'm out of here. You're leaving a world in computer science of like pure logic, right?

3:59At the end of the day, if there's a bug, you can always run it to ground. And that's because A, these are systems that run like clocks. B, we design them. We design them, right? And so like you at the end of the day can go in and figure it out. And in biology, it's like, sometimes you can, right? And sometimes it's just a part of the biology that we just frankly don't understand well enough that's broken and like tough luck. And you've got to like stomach that. And one of the things, my favorite things about AI is like, how does that neural network? Nobody knows. Like you're about to experience like the analysis of these neural nets is going to look like systems biology, right?

4:38It's going to be like, go in and like try to back, figure out a thing that you didn't design, my friends. And so like that, that'll be your first taste of really feeling like a biological engineer, right? But like, why bother? Like, why work with these neural nets that, by God, you actually can't easily debug and understand why it's hallucinating and all this stuff? And the answer is because they're powerful. It's worth it. And that's the same reason you want to do biological engineering. Like, even though it's unpredictable, even though it's going to be so frustrating, it's not going to do what you want, and blah, blah, blah.

5:08People are going to – like, it's because the substrate is incredible, right? It self-replicates. It self-assembles. We have nothing else like it in the physical world. So you want to work with it, even though it's hard. You know, that's what ultimately gets people passionate about this stuff. Yeah, it totally makes sense. And I think one can argue that neural networks are actually heading even more in that direction, because as people build systems that can code themselves, we're going to end up with evolutionary systems that are completely non-designed. And I think then we're truly in the world of biology, where you have evolution kicking in.

5:37And, you know, to your point, evolution is really messy, right? It's always optimizing for the utility of something versus the form of it. it reuses parts aggressively. It creates enormous redundancies in weird ways that you don't know that there's a perturbation here and it propagates across like in weird ways. And so I think people are really underestimating what happens once we have self-evolving neural nets, which I think is coming quite soon. Be still my heart. It's going to be great. It's so cool. I mean, that's so cool. It's so worth it. Yeah. I mean, it's neat because there's a magic to it, right?

6:08Like again, people like different stuff. And like, I get that. I think this would be like one of the big cultural divides. There's a certain kind of mind that really likes things to be predictable. There's some people who just like the magic, right? Like some of what's cool about biology is how hard it is to understand. So how does Ginkgo go about harvesting all this, you know, this shift in biology in terms of the ability to manipulate these systems using molecular biology and molecular cloning techniques and then software tooling and other things? Can you tell us a bit more about the company and where you focus and what you've done with it all to date?

6:37So one of the ideas that we tried to bring over from computer science was abstraction, right? And what is abstraction? Well, in Tom's era of computing, in order to be a computer scientist, you had to be an electrical engineer. Because how do you program a computer if you don't know how a computer works? Now, obviously today, like an eight-year-old is able to program a thing on their iPad by drawing boxes around. So like, what happened, right? Like, well, assembly language, operating system, programming language, graphical programming language. We build all these abstraction layers to split the disciplines of electrical engineering and computer science into their own paths, both of which had very long roads, right?

7:13And so one of the big things we did at Ginkgo, we started the company 15 years ago, unimaginably a long time here, was to do that split from the get-go. So we have part of our infrastructure, we call it a foundry, taking a page from SEMI's, that is basically a group whose whole job is automate and scale the lab work, okay? And move away from a system where that lab work is being done by hand by a scientist. And then the DNA programmers, who are really typically scientifically trained and PhD biology types, they order from that system to get their work done. That is actually very difficult to pull off.

7:47It's culturally difficult because like a scientist does not want somebody else to do their experiments and like they're a good scientist. You know, like there's a whole long laundry list of why it's hard. Not to mention that when you first try to build the infrastructure, it sucks. OK, and so the Foundry team has been able to drive enormous scale economics in doing the lab work, which gives us the data of lots of different genetic designs that we've tested, which is exactly going to be useful for the AI stuff. But it's also just generally useful, right, because you're you got to try a lot of designs to get the cell to do the thing you want it to do.

8:18And so that's that's been probably like one of the biggest activities the last decade at Ginkgo has been like driving that scale. Jason, what's the right way to think about the abstraction between like your customers and then your DNA programmers? Like what's the spec that gets passed over or how should we understand the science they do versus you? So today, the way that it works is basically a customer of Ginkgo's would be like, you know, like a recent customer is Merck. OK, like Merck, Nova Nordisk, Biogen on the pharma side, Bayer, Syngenta, Corteva, the biggest ad companies in the world, all customers, then a lot of startups.

8:48Right. And the way they interface with us is that we basically are agreeing on a spec. We're like, OK, here's what I would like to sell to do. They tell us. And we agree on like a timeline to develop it. And we're kind of like a prop software development shop. Like we're going to like make it for the customer and then license it to them. And they'll take and they'll own it to go develop their product. And in exchange for that, we'll get a royalty and also get some payment along the way. All right. That's the business model. Today, their scientists don't use our infrastructure. My scientists, or our scientists here at Ginkgo, they use the infrastructure and they have this interface with the customer about hitting goals.

9:24That's mostly a technical limitation. I think it would be very cool, ultimately, to have scientists at all these companies accessing our infrastructure directly. It's just too early. That's the problem. And then how does AI come into the picture? When and how did you start using it? And has this current wave of AI impacted you? Or how much do diffusion models, LLMs, et cetera, matter relative to what you're doing? So the short answer is like, we have you, we do a lot of protein engineering. So, you know, you want to program a bacteria, right? Okay. So bacteria has a 3 million letter genome and like a customer has asked you to express a protein.

10:01And remember a protein are basically like the little pieces of nanotechnology inside the cell that bump into each other and like do all the things like you're sitting there, you're like a big giant bag of proteins. Right. And so, and so like they want to make a lot of this protein because it's going to go into cold water laundry detergent okay so people don't realize this but like the reason cold water laundry detergent doesn't need hot water is because there's enzymes in there okay proteins and so they want to make a lot of this protein and by the way if they could make it more active like break up dirt faster like whatever reaction it is catalyzing so if you remember chemistry class like a catalyst makes a certain chemical reaction happen faster than if you don't have the catalyst okay so this enzyme is a catalyst and I want to make it also just better.

10:43So I want to improve the quality of the enzyme and I want to make a ton of it. All right, that's the spec. And so how do we do that today? Well, we would have, for example, a host strain that's really good at producing a lot of protein to begin with. Okay, so think of that more like an existing software library. So that's one form of leverage from existing data assets. It's literally like a hard physical asset, an actual microbe with a genome that I engineered in a project previously that is useful for your project. So now we're starting from a good place. We're already starting to make a lot of protein.

11:14But you want to make it more active. You want to make more of the catalysis. Okay, how do you do that? Well, remember, that protein is encoded in DNA, and the sequence of DNA determines effectively everything about that protein. But in this case, what you care about is how good of a catalyst is it. And so you go in there and you have certain tools to try to model the protein, this, that, and the other thing. And you make some choices and along with the software tools, and you say, I want to try these 1000 designs of the DNA in the lab and see how they do. And you then get that data back on how they perform.

11:52You use that to update these design tools you have, and you do it again. And that's what we've been doing at Ginkgo for a long time, including with neural nets and all this stuff, the latest and greatest, you know, all that. But just on that, like, data asset. I think the new idea is on the back, and this is, we just announced a deal with Google a couple weeks ago. The new idea is, can I make a foundation model that will be additive to what I've previously been doing just with the data I was getting on enzymes? And so now I have a foundation model that really is not specific to catalysis or anything else.

12:25it like speaks protein, you know, right? Like just like GPT-4 speaks English, right? That's what we're going to try out with Google. That's what we think is like a really new idea. And people, you know, there's people obviously work on it. And obviously Google themselves with AlphaFold was like one of the first generations of this. But we see a lot of ways to make it better and make it bigger and all the things. And so we'll see. We'll see how it goes. Let me just also give you like one other thing why I think bio is particularly interesting for folks that are interested in AI in general. So this whole idea of like a foundation model plus fine-tuning with specialized data, right?

12:58Like we all, like all the people that pay attention to AI like understand this idea, right? So let's just take it in one of the categories of English, like legal, right? Like LexisNexis, they have all this data. We're going to fine-tune GPT, all this stuff, okay. That thing has to compete with like a lawyer at Robes and Gray, right? And a lawyer at Robes and Gray has trained for 15 years, you know, being taught by other humans how to do law. They are writing contracts that were designed to be understood by human brains. They work the way we think. They're writing that contract in English, a language that co-evolved with our brains, you know, just language in general co-evolved with our brains to also give us leverage from how our brains work.

13:37And we're asking this computer brain, a neural net, to compete with us on our turf. But it's a pretty high bar that it's got to compete with. Now let's go over into biology. I remind you, it runs on code, sequential letters, feels a lot like language. It ain't our language, right? We did not invent it. We did not understand it. We did not speak it. We did not read it, write it. And so I feel like these computer brains are going to kick our ass a lot faster in this domain than they do in English, right? I think the applications in AI are all going to be like, replace the intern. not the partner at Rose and Gray at best for a while.

14:22Whereas over in bio, it could quickly become the best. If you're looking to understand where AI is really going to flip the script and not be kind of a low-level Clay Christensen-style disruption, which I think is sort of what's happening in English, but rather be a split in the atom, it's bio. When people talk about protein folding-related models, and to your point, there's things like AlphaFold And there's a few new companies that have been set up to basically focus on protein folding models because of the breakthroughs in AI. They kind of divide it into a few markets, right? There's sort of the pharma market, which is better designed for pharmaceuticals and biologics that are used as drugs.

15:01And then there's more of the industrial, the catalysts, the ag, those sorts of things. I'm sort of curious, like, how do you think about those relative markets just in terms of sheer market size? Because if you look at the cost of developing a drug, it's like 1.5 billion per drug. but very little of it actually goes into the underlying molecule that's being used, relative. Most of it's clinical trials. But for ag or for catalysts or for other things, a lot of it could actually go into the molecule. So I'm just sort of curious how you view those as relative markets for this kind of stuff. So Ginkgo has like an AWS style business model, right?

15:30Like you need compute. I don't really care if it's, you know, if you're AWS, whether it's medicine or they're a startup or you're video streaming, like rock and roll, right? So we have an attitude that like we're supporting cell engineering wherever it is. it is definitely different by market, like pretty substantially, right? Both the assets you need, the enterprise sales, everything is different. The biotech, you know, it's not a small industry, but it depends like which side of the house you're looking at. If it's fees, it's probably a little more the more valuable markets that you could get more research fees.

16:01But royalty is a different game, right? Like certain things go to market a lot faster. That's how I see it. But the real problem is like there aren't any platform services. So like the other thing that's just wild to people in the tech industry is like, where's the fast platforms? Like, where's all the horizontal stuff? Where's the operating systems? And it's like nowhere, like vertical integration, vertical integration, vertical integration, Merck, Pfizer, Bayer, Syngenta. Like every one of these companies is like its own tech stack top to bottom. You know, and the closest thing they do to anything vertical is buy equipment from the same people.

16:32But like, it's really fascinating, like totally different industry structure. And so that surprises people, right? Right. Jason, what's the most rational way to understand that? Because you look at that as a non-bio person and you're just, yeah, it doesn't make any sense, right? So the rational way is the work being done cross product is too dissimilar to support common platforms. I don't agree with this. Obviously, my entire business model is predicated on that statement being false. But that is the reason. You're like, oh, well, because that platform you build for customer A will never read on customer B.

17:09And so now you're just bad, right? Like you're having to build a new platform for every customer. You're getting, you know, leverage. You're getting none of the reasons that like operating systems and data centers, like why does data centers work? Because compute is really generic, right? And like you can use software to make it different, but the hardware underneath is all common. And now we're seeing a little edge case difference here, right? Like actually the CPUs aren't that useful for the AI. Now everyone's freaking out about the GPUs. So we're having like an instance of like hardware variability, but the argument would be that type of hardware variability you're seeing is sort of like per company, okay, right?

17:39Like, you know, or at least per modality in pharma, which modality is like a fancy word for type of drug, right? Like gene therapy is going to have very different stuff than, you know, whatever. And there's truth to that, right? Like it's not, that's not a false statement. It's just a question of degree. And we happen to believe that when it comes to the engineering of organisms, that that is at parts common. But plenty of people think we're wrong. Yeah, it seems like if you just go back to this analogy of like human design computation, where you're building systems from the ground up that you can understand, but you're all discovering the same set of systems with common building blocks and the need for data analysis, it would shock me if that would not eventually be true.

18:19And it's like a temporal cultural figment of these companies. Okay, maybe one more general question about how to think about AI in pharma writ large. Why do you think we haven't seen AI discover drugs yet? Because people have been talking about it for a long time. Will we and will we see it soon? Well, so first off, I would say like the people that talk about for a long time is sort of like saying three years ago, like, why haven't we seen good natural language processing in AI? People have been talking about it forever. Right. So I think there is an element of like you need the breakthrough. Right.

18:53How's the neural net's been big enough? The big limitation in bio is the availability of data to train these things. Right. And so you have this tough situation where like everyone is doing these models of training on the same data. Right. And so one of our advantages is because we have a ton of data. That's a real gap that I think is I think partially it's like have people gone big enough for it to have happened yet? And I think I think now people are trying like we're going to try recursion is another great example. And like, yeah, it might still not be big enough or more likely it's not enough data.

19:20Right. And like there's nothing stopping you from making a giant neural net at this point. You know, like the tech industry is going to commoditize that infrastructure. But like you might not have enough data to give it to solve the problem you're asking, Sarah. Right. Where does Ginkgo's data come from? Is it like your own experimental data? Yeah, we have a 300 ,000 square foot robotic lab that they built in the last 10 years. And so we generate that and we do it in service of our customer projects. We can do our own data generation. But yeah, that's where it comes from. I want to talk a little bit about one area that Ginkgo has been an expert in, which is infectious disease.

19:54Right. Right. Can you talk a little bit about the work you guys do here? And I think a question everybody cares about is like, are we prepared for another global pandemic? What has changed since COVID-19? Yeah, I think like the reality is infectious disease is really scary and bad. Right. But like the big the big lesson of COVID is modern health care systems and our current infrastructure does not render us immune, even in the developed world from pandemic scale, infectious disease. Period. We don't just allow ourselves to not have defenses against things that are like society killers that are known to exist.

20:32You know, right? Like, this is not like, you know, a fanciful idea. Like, it freaking happened two years ago. Right? You know? And so what should we build? Right? And the answer is, like, a lot of different things. Right? Like, we should build rapid vaccine response, which is really good, through Operation Warp Speed and kind of what we figured out with mRNA vaccines. And just, like, I got a target. I got vaccines for the entire country in three months. Not every version of this thing is vaccinatable. The other one we've been big believers in is like monitoring, like radar. I grew up in Florida, right?

21:01Like we have radar systems that warn us for hurricanes. Okay. My co-founder, Tom Knight, obviously a lot older than me, was explaining to me that when he was a kid, they would get three hours warning for a hurricane. Currently, we find out about hurricane COVID after it has landed in New York City a week ago. Okay. like unacceptable, right? So one of the things we're doing is like with the CDC, we run programs, we collect wastewater from inbound airplanes and we sequence the DNA and we look for pathogens. We monitor variants and all this, both for flu and COVID. And I can add other things to that list.

21:38We have a similar program actually in Doha airport in Qatar. We've got a program in Ukraine. And think of these like bio radar stations. And that's get you baseline because you also want to, you will look for anomalies, right? So like that whole thing has been missing. And so we think that's like where you start. And then you want rapid response. Like basically patch. Think like cybersecurity. Like it should feel like that's your answer. Right. Like like cybersecurity, I think is a bit the mental model for like what the future of infectious disease response looks like. Persistent monitoring, rapid response, kill it.

22:08And the beautiful thing is and it's scary and beautiful. Remember, these things replicate. So if you can snuff it out at the beginning, you win. Like speed will matter. Right. I'm chair of a national security commission down in D.C. on emerging biotech. And like the DOD just put out their biosecurity. It's like basically biodefense posture review. And like the DOD maintains like millisecond preparedness in this country. Right. Like I know that seems crazy, but like that's kind of like how you ought to treat these things. Yeah, you actually saw that with SARS, right? Because SARS, they both snuffed out with the original form, but then it leaked four times in the first two years after it was cultured in a lab.