#171 Anna Marie Wagner: Harnessing AI for Synthetic Biology

0:00So customer comes to us, they've got a spec, and again, this is across all different industries. So it's as varied as I want my corn to fertilize itself, you know, and I'm interested in engineering a microbe that will live inside the corn and fix nitrogen out of the air. And so I can reduce my nitrogen addition by 50 percent while not harming plant yield. Like that could be the spec or the spec could be, again, something as specific as I said, catalyzing a particular reaction. Or it could be I want to discover a new RNA therapeutic. Hi, I'm Craig Smith, and this is Eye on AI. In this episode, we explore the dynamic world of synthetic biology with Anna Marie Wagner, head of AI at Ginkgo Bioworks.

0:48Anna Marie shares how their platform leverages AI to push the boundaries of biotechnology, touching on projects like self-fertilizing corn and the quest for biodegradable materials. We delve into Ginkgo's collaboration with Google, the nuances of building foundation models for biological data, and the broader implications for biosecurity and sustainability. The conversation provides a window into a future where biology and machine learning converge to innovate and heal. I hope you find the conversation as fascinating as I did. Introduce yourself, Anna Marie. tell us a little bit about who you are and where you came from and what you do.

1:37Sure. Thanks for having me on. So I have been at Ginkgo for about five years. I wear a number of different hats. I am our head of corporate development. And so in that role, I oversee M &A, licensing strategic partnerships. And that was the role I joined to lead. But as part of that, I negotiated our collaboration with Google, which I'm sure we'll talk about later, and started working very closely with all of our scientists across the business that are deploying AI across many different disciplines, and got candidly just really excited about everything that we were doing in AI and the potential for AI in our field, and so took over a new role about five months ago or so as our head of AI, overseeing our AI strategy and commercialization efforts, so really working with our customers.

2:24yeah and so tell us uh about ginkgo bioworks and and what what you guys are focused on um i mean that's it's there's so much happening in in biotech and ai yeah absolutely yes it might it might be helpful if i give you just a little bit of the history not just of the company but of the backgrounds of our founders because it does help elucidate some of the philosophy by which we've approached the business. So Ginkgo itself was started about 15 years ago, five co-founders, all of whom are still with the business now 15 years later, but they've actually been working with each other for over 20 years.

3:02So they all met at MIT. Four of them were doing PhD programs in the newly created sort of bioengineering field or synthetic biology field. And the fifth was a professor. His name is Tom Knight. He had been at MIT for 40 odd years, originally in the electrical engineering department. So he built one of the early semiconductors, worked on LISP, built the first real-time debugger. Eventually, as semiconductors came onto the scene, he taught the semiconductor design class at MIT. And it was around the early 90s he got interested in biology for a couple of reasons. One is as a substrate for nanoscale precision, which as a semiconductor designer was something he thought a lot about as he thought about the future of Moore's law but also because biology is a programmable physical discipline and you know so it runs on code ACTG not zeros and ones but we could read that code right with DNA sequencing we can write that code with DNA synthesis and he's like we should be able to program this stuff And so he's approaching the field like, well, what would it mean to program the physical world?

4:18That's pretty neat. And so to his great credit, he's in his mid-40s. He starts taking undergrad biology classes. He builds a wet lab in the computer science department, which, as you can imagine, the rest of the computer science department was really thrilled about. And he started assembling all of these engineers from different disciplines, you know, mechanical engineering, chemical engineering, computer science, electrical engineering, obviously, to start thinking about what would it mean to make biology an engineering discipline rather than just this artistic field of discovering nature. It's a very different way of thinking about the field.

4:56And there were a few observations relatively early on. One was, you know, something we needed to build, first of all, was like the layers of abstraction, right? So to be a programmer of biology, the first thing you were told to do is pick up a pipette. It was sort of like being a programmer of computers in the 60s. When to program a computer, you needed to understand the electronics, not just programming languages didn't exist, right? And so we are still living in that era of programming biology, where to be a programmer, you had to understand the fundamental mechanics of the machine. And so he's like, okay, how do we build those layers of abstraction?

5:36That was kind of one key question. Another key observation was because this is a physical discipline, physical disciplines tend to benefit from scale economics. And so how do you have scale economics and biology? You're going to invest in the sorts of things that drove scale economics and things like industrial manufacturing, right? So automation, miniaturization, things like that. And that led to, I think, business model decisions ultimately, you know, 15 years later when Ginkgo got started around being a platform that was going to serve many different areas of biological design rather than focusing on a single product, which wouldn't allow us to make a the kinds of investments at scale that we knew were going to be needed in order to actually drop down the cost of doing this engineering work sufficiently to make it scalable.

6:27And then the third piece and final piece that he really noticed or question he asked at the time was, well, where are the code libraries? I'm going to program some biological code. Where's the library? Like we have this stuff in computer science. Where is it in biology? Didn't exist. All of the code that had been created was happening in silos in small biotech companies that were making their individual drug. And they're religiously protective of it because the only thing that creates value in biotech is IP. So everyone is protecting their IP. And so there was a ton of recreating the wheel that was happening in the biotech field.

7:01And so that, again, led Ginkgo to make a very different kind of business model decision, which is that when we create IP, we want that IP to be shared broadly. We know that if we engineer a protein that's useful for one domain, it's very likely useful in many, many other domains. And we've seen that time and time again in the history of Ginkgo. But that was quite countercultural in the field. And so I just want to give that history because it helps explain what Ginkgo is today, which is a horizontal platform with a very large, 300 ,000 square feet or so of infrastructure, you know, largely automated labs that is acting as a platform to enable R &D to happen more efficiently, faster, ideally higher probability of success for customers across many, many different industries.

7:50And then the business model alongside that is we get paid sort of like AWS, you know, service fees for doing the work on the platform. And then more like an Apple App Store or something, we're getting paid on success. So if the customer's program actually works and they sell product, we get a royalty on that. And what kinds of products? Because is this drug discovery? Is this, you know, engine, gene editing of plants? Is it, yeah, sort of give me a scope of what you guys covered. Yes. So yes, all of the above. So sort of counterintuitively, we got our start outside of pharma, even though it's obvious to everybody that pharma is the biggest market for biotech.

8:33Had we walked into Pfizer 15 years ago and, you know, we were like five people and some robots, you know, they would have laughed us out of the room. They were way better at doing biotech R &D than we were 15 years ago. Right. And so we started in areas like flavors and fragrances, you know, clearly a biological product. Right. These are specialty compounds that are extracted from flowers and fruits and things like that. But the companies that were making these things had no real biotech expertise. They had supply chain problems, they had cost of goods problems, and they didn't really know how to solve those.

9:07And so it was a nice interface between our capability set and our partner's capability set. And so that was where we had our first product. So things like companies like Roburte and Givadon, you know, very large European flavors and fragrances companies where we got our start. And then, you know, sort of slowly we marched up the credibility curve where we then started working in more and more sophisticated end markets because we had the data to show. all right, we successfully delivered on products in this area, programs in this area. Let's go tackle the next area. So specialty chemicals, food, agriculture, ultimately pharma.

9:42Today, our business is split about a third, a third, a third between biopharma. So customers like Pfizer, Merck, Nova Nordisk, Biogen, all customers of ours across many different modalities in biopharma and everything from discovery to manufacturing. And then, you know, in the chemicals industry, We work with Sumitomo. We work with Solve, some of the big chemicals companies. And then in agriculture, we work with Bayer, Syngenta, and Corteva, three of the biggest ag companies in the world, both on things like traits, but also on things like microbes that live in the soil and help fertilize your plants for you in the soil.

10:21So really interesting variety of work that is running on a common platform, which, again, is kind of unheard of in biotech land. Yeah, that's fascinating. And just to clarify my understanding, because this is way out of my area of knowledge, in pharma, there's a lot of talk right now about small molecules as opposed to biologics. Are you guys focused on biologics or do small molecules, is that part of the mix? We do everything. So yeah, we have small molecule programs and we have sort of complex, so everything from small molecules to large molecules to complex therapies like cell and gene therapies, we have programs in.

11:12On the small molecule side, typically where we have a particular advantage is when you're looking at what are called natural products. So these are small molecules that have otherwise been created in nature. So, for example, a lot of our antibiotics come from microbes that live in the soil because all these little bacterias are fighting each other. And so if you want to fight that bacteria, it turns out some other bacteria is probably already successfully fought it away. And so we found a lot of that chemistry by looking at biology. And what's interesting is that quite often these sort of magical molecules are quite hard to synthesize chemically.

11:48They're quite complex chemistries. And so even though they're small molecules, making them with traditional chemical synthesis pathways is hard. And so you also need biology to produce them. So that's one version. And then more on the manufacturing side, similarly, so for example, Merck is one of our customers in this space. They're producing small molecules. They already sell these drugs. Right now, they might have some really expensive or messy or inefficient chemical synthesis stuff. And so one of the project areas we work on is something called biocatalysis, which is can you do a chemical synthesis step using a protein, using an enzyme?

12:29And so we would discover and engineer an enzyme to catalyze that chemical reaction more efficiently than the current chemical synthesis pathway that they're using. So those are a couple examples of small molecule products. But the industry writ large is quite focused on biologics and cell and gene therapy as well. Yeah. Where does AI come in to the process, to your workflow? So, again, I'll give you a little bit of history. Ginkgo, historically, I'd say for the first 14 years of our 15 year history, Ginkgo was sort of accused of being this like brute force experimentalist. Like, we're just going to solve all the biological problems by throwing scale at it.

13:11And remember, biology is this, like, artistic discipline. So, you know, really good biologists sort of looked down their nose at this brute force scale that we were applying. And to be honest, the way that Ginkgo originally thought about data, and remember that those code libraries, was around the successful experiments. Like, we're going to reuse debugged code that we know works. We're going to reuse those across applications. But as it turns out, the real magic of operating at the scale that we have operated at and that we've built is that we are not just studying the success stories. We're also collecting a whole lot of data on the 99 % of experiments that did not generate success.

13:53And as you well know, when you're thinking about training a model, you want to have, yes, examples of what good looks like, but you also want to have a hell of a lot of examples about what bad looks like. And you can learn just as much from that, if not much more, just because there's candidly a lot more failures than successes in biology. And so now as we think about what is our role in AI as applied to biology, the rest of the industry really hasn't been designed or set up in a way to collect that kind of data. You know, real labeled data on how biology functions at scale. Like that is a pretty unique infrastructure that we've built.

14:34So to give you a couple of tactical examples, our protein engineering team, big users of AI, right? So we've deployed several AI models inside Ginkgo, sort of homegrown, and we, you know, we'll use anything that's out there. We're sort of ambivalent as to what the architecture is, because the thing that we really bring to bear is data, right? We have a much, much larger genome collection than is publicly available. Our proprietary metagenomic collection is about 10 times the size as the public databases. And then we have all this labeled data on how those proteins actually function in the lab against a battery of tests.

15:10And so now when we're designing a protein for a customer, we can deploy all that data into these models and fine tune them to yield better results for the customers. So protein engineering is a big area for us. DNA design is another big area for us. So, for example, trying to improve the expression of a particular gene sequence by engineering promoters and other kind of non-coding regions of DNA. Those are kind of easy use cases for us today. But the reality is anywhere we're generating data, that data is useful for training increasingly complex models over time. And so you can certainly imagine this getting to a point where we're creating models about how an entire cell functions.

15:54Right now we're starting with the building blocks, proteins, DNA, RNA. Then you get into things like systems biology and pathways and things like that. Then you get into broader cellular function. Then you could imagine trying to predict how entire ecosystems behave together. There's an immense amount of complexity that is beyond the scope of what anyone can really do right now. But even the advances we're seeing just on those building blocks are quite remarkable. And so the AI that you're working with is primarily a search function. You're looking for attributes in a database of molecules, or are you using it to generate new molecules that have certain properties?

16:44Yeah, so it'll be both. So our projects, so again, let's take one of those enzyme engineering projects that would typically start with an ML guided search campaign of the metagenomic landscape that we have to identify interesting starting points. And then we would use those starting points, generate data about them, and then in a kind of a closed loop system, generate new designs that we think are interesting and what's important about the way that we operate, again, because we have the labs and we can actually generate this data at scale in a pretty quick way, is rather than on the first round saying, all right, we got one shot, let's make the best thing we can make, we can really design the experiments to maximize learning.

17:30And typically, we're optimizing across many different parameters at the same time. So we will generatively design new proteins, right, which are going to be based on a scaffold that we've found in our databases, you know, from an existing piece of biology. And that's useful because you want to, you don't want to, like, there's a reason all that biology evolved. And usually it evolved because it functions well, it grows well it's soluble like they're going to be a bunch of features of the biology that's useful and you want to preserve um so we'll start with that emma guided search we'll then generate a library that maximizes our ability to learn across the many different parameters we care about so it might be uh activity it might be selectivity um so if you are off target you know kind of interactions um and then we would design subsequent libraries to maximize across those dimensions in parallel But those are also using generative models to create new protein designs in that case.

18:27Yeah. Do you know the company Insilico, which is a drug discovery company? They're using AI to generate or discover new molecules, small molecules. But they also have an automated wet lab in which they can synthesize these molecules or at least samples. And then they send it out to a larger lab for production. Is that similar to what you guys are doing other than the fact that they're focused on developing drugs? And you guys are really a platform for anybody to use? So I think there's a lot of really great research that's happening across that intersection of AI and biology. I think what's interesting about Ginkgo is that, again, we're quite ambivalent as to what architectures are going to end up working best.

19:27And we tend to be real beneficiaries of that kind of innovation that's happening elsewhere in the industry. And interestingly, I honestly see more opportunities to partner than to compete, right? And so we can, in many ways, operate in a couple different venues with some of these AI-focused companies. One is to help their platform reach more customers. We just integrate the model. Today, our protein engineering team uses, I don't know, seven or eight different models in all of our campaigns. So we'll integrate different models into our work, and we'll lean on the ones that then start yielding the best results.

20:06And we're obviously building our own internal models, but there's benefit to having that diversity. And then the second is, you know, as you've seen in the large language model space with human language specifically, there are companies that have specialized in building architectures and companies that have specialized in building data. And so take, for example, OpenAI, building GPT-4, but then ScaleAI, which created a lot of training data and allowed them to do reinforcement learning so that you got ChatGPT out of it. One way to think about Ginkgo is that we're like a ScaleAI for the broader kind of biotech ecosystem as it relates to AI.

20:45We're creating that data that's going to make all of these models better. And so our goal is to get a customer result. We don't have our own pipeline or anything else. And so we're very happy to use the best of what other folks are developing. And where folks have labs, they tend to be quite narrowly focused on the questions that they're interested in for their own model or their own pipeline. And so there do tend to be still really great collaboration areas to round out the type of data that you would want, or to the point you made earlier, once you start thinking about scaling this up, making it more industrial, wanting to really manufacture this stuff, how do you do that?

21:20And can you walk me through an interesting, because I'm sure you have many case studies at hand, an interesting case study of sort of what happens with the client when they come to you guys, what's the process, and then the workflow and the work product as it is. Yeah, it is quite interesting because we're often compared to a contract research organization, a CRO, and in many ways, we are like a CRO. But traditionally, the way a customer would interact with a CRO is they would say, I need this study done. Can you please do this study for me? It's an outsourced service, but the customer knows the work that they want to get done.

22:11When customers come to Ginkgo, certainly we're providing them a lot of services, but they're not dictating, all right, I want you to run this assay on your mass specs, and I want you to run whatever, this fermentation process in your labs, etc. They're coming to us with a problem. I need that, I need to replace this chemical step with an enzyme, and it needs to cost less than this, it needs to catalyze at this efficiency, it needs to have this level of specificity, like they're giving us specs of a product. They're not giving us a research plan. We develop the research plan. And it quite often looks very different from what the customer either might have tried internally or what they would be able to do anywhere else because of the scale of the platform we've built.

22:52We're able to try a lot more breadth. And quite often what we see is you're trading a local optima for a global optima, right? When you can only afford to search a relatively narrow space, you have to make pretty safe bets. You can't afford to try off-the-wall things. When you can, you know, increase the scale you're operating at by three, four, five, six times, sorry, the order of magnitude by three, four, five, six times, then you can suddenly start to ask much more interesting questions. And we very often see that the winning design for a particular program might look absolutely nothing like their best, their best design coming into the program.

23:37And so, yeah, so customer comes to us, they've got a spec. And again, this, this is across all different industries. So it's as varied as I want my corn to fertilize itself, you know, and, and I'm, I'm interested in engineering a microbe that will live inside the corn and fix nitrogen out of the air. And so I can reduce my nitrogen addition by 50 % while not, you know, not harming plant yield. Like that could be the spec or the spec could be, again, something as specific as I said, catalyzing a particular reaction. Or it could be I want to design, I want to discover a new RNA therapeutic. I don't even know what the right design should be.

24:16I don't know if it should be circular. I don't know if it should be linear. I don't know how to deliver it. You know, you've got to work on a whole package for some of these projects. And the customer has got a product idea in mind. And then it's our scientist's job to bring together the pieces of technology that are necessary to deliver on that. Yeah. Actually, of those examples, the corn is the easiest one for a lay person to understand. Was that a real example? That's a real program we're working on with Bayer. Yeah. Yeah. Can you – so how do you start and what's the process that you go through?

24:55Yeah. So that is it. that is a real moonshot problem. And this is a like, you know, it's like a$70 billion a year industry in nitrogen fertilizer. So it's a really cool problem to work on. It's also, by the way, fun fact, nitrogen fertilizer alone is something like 5 % each of global energy consumption and global greenhouse gas production. So I mean, this is like, if you can reduce the amount of nitrogen fertilizer that is needed to feed our planet, like this is hugely, hugely, hugely valuable. Okay, so where would you start? So there's sort of two components to that. This is a drastic oversimplification, but simplistically, there are two components to this challenge.

25:36One component is how do I get the bug to live happily inside corn or wheat or rice or, you know, you name it, your cereal crop of choice. And so there you're engineering, you're trying to find bugs that live, And there already are, by the way, lots of little microbes that live inside plants, very happily. And then the second key question is how do you get that bug to fix nitrogen? Well, it turns out there are bugs in the world that do that naturally. Like you don't need to fertilize soy. You might be like crop rotation. You might remember from like middle school, we used to rotate crops. And the reason you rotate crops is some of those crops, the legumes, they fix nitrogen.

26:15They replenish the soil with all of those nutrients because they've got these little microbes that live in their root structures that fix nitrogen. They do Haberbosch, basically. And so now the question is, can I figure out the engineering circuitry of the bugs that do Haberbosch? And can I put that circuitry inside the types of bugs that are very happy living inside corn or inside rice? and so you're you're and you I suppose you could do it either way but like you could try to make those bugs be happier with corn you know they're they're obviously many different ways to to solve the type of problem but you're you're figuring out what is the what is the DNA that codes for Habermasch basically and then you're trying to figure out what is the DNA that codes for corn lovingness and you're trying to get both of those things to exist in the same bug and then you've got all sorts of things like how do you deliver it is it a seed treatment is it a soil additive you know like all sorts of other other things but um that would be the the basics of that type of a program right so so uh let's take the the which sounds simpler uh or maybe not but uh figuring out the uh the genetic mechanism for uh fixing nitrogen in the soil yeah walk me through how theoretically how you would do that yeah yeah so it's a it's a metabolic pathway right so the the microbe is eating something in the air right it's like what's in the air there's some nitrogen floating around, there's some oxygen floating around, there's some carbon floating around, right?

28:00And these bugs are ingesting those molecules. And then there are a series of proteins that are coded in the DNA, and those proteins catalyze certain chemical reactions. And so many of those chemical reactions, we sort of understand. Like, we understand generally, well, first, you know, all right, the plant eats the carbon dioxide, and then the carbon, sorry, I'm not a biologist anymore. I don't even remember these. The carbon dioxide is broken down into something else. We do understand these metabolic pathways in general. And so then we can start identifying the proteins that are responsible for catalyzing those reactions.

28:36And then it's a question of, so you generally have a decent starting point as to what the set of chemical reactions needs to be. Sometimes, by the way, you can find more efficient pathways. So like, hey, what if I took out these two steps and instead just did that one that kind of circumnavigated the chemical, chemical steps. But typically, you'll find, okay, there are, I don't know, let's say four steps involved, making this up, four steps involved in converting, you know, carbon dioxide plus, you know, whatever else into ammonia, effectively. Here are what those steps are. Here are the proteins that code for those in the microbes that live near soy.

29:18Then there's a question of like, you know, load balancing, basically, like how much of these things do I need to get expressed? And so that's where understanding things like promoters, which are like non-coding regions of DNA matters, because you need to make sure the right amounts of these proteins are getting expressed in your new bug. You need to make sure your new bug isn't getting killed because of these new pieces of DNA that you're adding or any of the intermediate molecules that are floating around in the cell now that are being catalyzed. So the hard part is not typically, let me figure out the proteins that are responsible for this.