Ginkgo Bioworks Holdings, Inc. (DNA) Earnings Call Transcript & Summary

Good morning, everyone. This is Steve Mah on Tools and Diagnostics team. My pleasure to introduce Jason Kelly, CEO from Ginkgo Bioworks to the Cowen Healthcare Conference. As a reminder, if you have any questions, please e-mail them to me at [email protected]. So Jason, I mean, I've known you many years, but do you want to give a quick intro to folks that are not know familiar with Ginkgo? Try to level set us.

Yes. Very happy to do that. And everyone can hear me okay, that sounds a little echoey up here. Here we go. How is that? Better? Okay. Good. Sure. Yes. So I'll kind of give a little bit of like a health care framing for what Ginkgo does. And so the easy way to think about what we're trying to pull off at Ginkgo is we're trying to centralize an activity that today is done in a distributed fashion. So then this is -- think of it like the discovery activities associated with biotech drugs, right? So about half of your new FDA-approved drugs are coming out of biotechnology, right? In my view, they have something in common, which is essentially -- each one is basically putting a designed piece of DNA into a cell and then putting that cell into a particular environment that makes this drug, right? So if you rewind the clock to 1982, that's putting a gene for human insulin into an e-coli cell into a metal tank in South San Francisco at Genentech, right? Go a few years forward, right? It's monoclonal antibody DNA inside a [ Cosell ] at ABBV, okay, making monoclonal antibodies, right? Cell therapy, right, Arie Belldegrun on our Board at Ginkgo, right? This is a piece of DNA inside of a T cell, okay, going into a patient with cancer, all right? Gene therapy, right, CRISPR Therapeutics announced a partnership with Vertex also on our Board, Reshma Kewalramani from Vertex to go after sickle cell, right? That's adding a piece of DNA into a patient cell in vivo. And then everyone knows about mRNA vaccines and Moderna, right, again, mRNA being delivered in VIVO. And my view on this is all those activities, the process leading up to when you put something in the clinic, is essentially cycling through genetic designs, putting them into cells, testing how they perform and then based on what you learn iterating on those designs. And the entire bed of Ginkgo is that, that activity, which looks really different across all those modalities is actually the same activity. And you can centralize that activity on top of automated facilities. And if you want, we're right down here in the Seaport but 10 minutes away from here, about a 250,000 square foot automated lab. We can talk about the technology, Steve. But the idea is that work today is done in a distributed fashion across thousands of biotech companies and large biopharma and we believe it should be centralized into a small number of automated facilities. And now that seems crazy, right? All these companies already have their own lab setups, they have scientists and lab coats Holding pipe heads doing this research. How can you drive centralization? And my argument would be if you look in the tech industry in the late '90s, early 2000s, everyone had an IT department everyone had servers and big setups in-house with those on-prem physical infrastructure and teams and along comes the cloud with a physical scale economic and data centers. And over time, that just wins, right? People move and they have all these excuses, they don't want to move. It's my data, it's this. I don't trust you, Amazon, whatever. But eventually, the scale economic of the centralized infrastructure wins. And that's the core bed of Ginkgo is that we're finally at the point we can talk again about the technology, where a centralized infrastructure to do preclinical biotech drug development is going to win over distributed infrastructure. So that's the short answer. We can talk about the tack and we can talk about the business model and see about how we're pursuing that. But you either believe that or you don't. If you believe that, I think Ginkgo's a good bet. If you think it deserves to be distributed for some reason, then we're not.

Yes. Yes, why don't we talk about the business model. So obviously, this is a health care conference, but Ginkgo does a tremendous amount of things in like the applied markets, right, whether that's flavors and fragrances, at bio, food, whatnot. Maybe just sticking, well, maybe you just say a little bit about that and opportunities there, and then I'll kind of dig into therapeutics. Can you maybe just kind of level set people on the how the technology can be used for what literally almost anything that's carbon-based or biology-based really, which is potentially everything except maybe oil drilling?

Yes. Anything involving a cell. Yes, that's the short answer. Yes. So again, what's the right business. Let's say we're right and then you want to centralize this activity, okay? So today, we have customers like Biogen, Novo Nordisk, Merck, right? They have their own research labs. Why have they outsourced work to us, okay? We have to have something that they don't have, okay? And so the 2 things are physical scale in automation, okay? And the reason this is important is the bigger your automation -- think of it like a factory for making chips or cars, the lower the unit economic of the thing you're making. And in the case of Ginkgo, what comes out of our automation are characterized genetic design. So if you want to cycle through, as an example, at the SITC conference, immunotherapy conference a few months back, we presented data that we had done in CAR-T, right, where we made these huge genetic libraries, 10,000 libraries tested in vivo, that scale of being able to generate those libraries, you just can't do unless you have highly automated mammalian genetic engineering running at a bigger scale than your scientists could do with the PIPE. So that's the first asset. Automation to allow you to try more genetic designs per research dollar. The second asset is data. So as we do all that work in the lab, we do it in a systematic way. We collect the data, and Steve, we should talk about this, but we have IP arrangements with our customers that Ginkgo retains the rights to reuse the IP. And so the second reason you work with us is you want access to our genetics to our data, to our learnings from projects we've done before. We have a project with Selecta and capsid design. -- that we have now a great capture library. If folks are doing AAV Gentherm, they want access to some interesting capsids to work on the problem of tropism, call us up. okay? They're available as part of our data assets in our service offering. So those are the 2 things, the automation and the services -- or sorry, the data or the reason people contract with us. All right. So then what's the business model? Very, very fancy CRO is the easy way to think about it with a pharma hat on, okay? Folks are contracting research to us. They're paying us a fee to do the work over the period of a project might be from like 1 to 2 years of research fees. And then the big difference between us and a traditional CRO, like someone doing animal trials like a Charles River or a small molecule, libraries like WuXi, is that we, number one, take a royalty or take milestones or equity, some type of value share on the product, on the drug product that is going to get produced with our platform. That's different than a traditional CRO. And then, number 2 is this IP arrangement I mentioned earlier. We're retaining rights to reuse the IP for other projects, also different than a traditional CRO. But other than that, it's a research services business, right? And we got now 96 active projects. And to answer your question, Steve, about 34% are pharma.

In pharma?

Yes, right. So -- and then about 1/3 are in ag and food. And then the remainder are in industrials, so things like chemicals and nutrition and flavors and fragrances, things like that.

Okay. Yes. Maybe let's talk about the value proposition of each of those things, right? Obviously, you talked about in your last earnings call, the downstream value for therapeutics is arguably higher than some of these applied markets. Maybe talk about the mix and you expect to keep about 1/3 of your new program adds and therapeutics? Or do you expect that to grow?

So a good question we get all the time is like, hey, it seems like biopharma is a much bigger research market for biotech than, say, fragrances. That would be true. It would seem like the value of a therapeutic drug is much greater than the value of a nutritional ingredient, also true. Okay. So why didn't we just start offering our services business in the best market for it, biopharma when we [ speedout ] -- we started the company now 15 years ago, but I'd say, over the last 10 years, why not? And the answer was I have to convince you to not use your internal team, but rather to outsource to me in order to get you as a customer of Ginkgo. Like I am not a product company. I don't have my own pipeline of drugs. I can't just think I'm so smart and my technology is so good that I can develop a drug. I have to sell a customer on it. And if I were to compare the delta between internal R&D at a fragrance company in biotech and Ginkgo's platform 8 years ago, we were better than them. They are a bunch of chemists. They're not biotech people, right? They weren't any good at biotech. And so I had to convince that biotech mattered in their industry, but I was better than them at it. Eight years ago, I could not have walked in to Biogen. I could not have walked in to Novo Nordisk incredibly said, my platform was better than theirs. Why? It was a lot smaller back then, right? I'm running a scale business, right? As our infrastructure gets bigger, it improves with scale like a data center or a chip fab. And so it's only been in the last, say, 4 years or so, Steve, that we've gotten into biopharma at all, right? And that is a function of the improvement in the technology. So now where will I go? Oh, you better believe more biopharma because now we proved it, right? Like we are signing up again, this is not my opinion. This is the opinion of our customers that we can win deals in that area that we have something to offer proprietary and it's such a better market than some of the places we got started in. So yes, I think you should expect us to do a lot more biopharma over the next 2 to 3 years. And then who knows, right? I think the wildcard in all this is that the technology gets substantially better as the difference between what you can do with an automated facility and at the lab bench improves maybe new markets open up. Yes.

Yes. Okay. Let's talk about the scale of your founders. I know you just opened Bioworks7, I spend a little bit of time on that. And is that -- is that sort of driving a lot of new partnerships?

Yes. So the -- yes, I'd say the 2 big things that drive our partnerships are -- well, let me say this, Biopharma, in general, is like very data-hungry in terms of choosing to work with us. So -- we like to give tours of our facilities. We can show the total output of our genetic engineering and mammalian but like what's really carried today is as we've been doing programs over the last 2.5 years. We now have interesting data that they see, okay, you can build whatever, that many CAR-T designs. You can build that much in AAVs. You can do this much in microbiome. And we can just show them the scale of the data. It's not on exactly their drug, obviously, it was with somebody else on something different, but it's in their modality. And that's the biggest driver. On the -- it's like the proof in the pudding data with something close to what they're interested in, not exactly what they're interested in, but close. The other thing is we did start -- we acquired redate acquisitions last year. So we acquired a, for example, a company called Circulares. Okay. It's a circular RNA company. It's not a Orna, okay, right? It's not like famous ones, right? But it doesn't have much of a pipeline of its own but it had really good technology and also had great technology around promoter screening. It's like a reusable asset in the genetics. And that has quickly turned into some great conversations around new deals in circular RNA. And so that type of asset is also useful for in our hands, more than it's useful in a small biotech because we have the sales relationships. We can put it on top of our automation and quickly demonstrate interesting new things. So there's also that as part of the sales process, too, so unique IP.

Okay. Yes. Got it. And how do you get your partners to sign off that you can use as IT? I mean, most biopharma is pretty secret about this.

Yes, sure enough Yes. So not an easy discussion, I would say, for starters. So our basic argument to the industry is that the isolation of genetic componentry across thousands of companies is reducing the odds of success collectively of getting drugs to market. okay? Like if someone develops a better capsid over here, that could be -- that is not the whole drug, right? That's a piece of the drug? Like what is the payload is a whole separate thing to innovate on. So why if you are over here coming up with some great ideas for payload, do you have to use a substandard capsid for your application just because somebody else with totally different idea, they're the ones who happen to invest the capsid. This is a joke, right? Like if you look at how the software industry operates as a comparative example, enormous amount of exchange of the functional componentry, like no one owns the if-then statement. No one owns a 4-loop in software. We've agreed there are certain low-level functions that are better off exchange. And then what you are selling is the higher-level combinations of those functions in proprietary pieces of code. This is the direction genetic medicine needs to move to or we're going to be stuck in this world where the genetic medicines are lot embarrassingly simple, right, like 1 or 2 genes, almost no regulation. These things are a joke, right? Like if you want to be able to unlock the true potential of biology, look at the complexity of a T cell we should be hitting a lot more in how that machine works with much more complicated genetic perturbations, but we're not going to do that if all of the assets are split among 100 different T cell companies. okay? And so what I'm creating is a walled garden in which everyone gets to share the reuse of the IP just like if you logged into Apple's ecosystem for the iPhone and they would give you a bunch of functions that make it a hell of a lot faster to develop an app. That's going to be the same thing if you're on top of Ginkgo's platform. It's going to be easier to develop genetic applications because I'm driving an IP regime that is not typical in the wider industry. And so we're very aggressive about that. We hold the line on that. And if people don't like it, then go back to your guy at the lab bench with a pipe [ head ], ultra low throughput, right, and try to compete.

Yes. No, that's a great analogy. Staying on the foundry and the capacity. So your business model is pretty dependent on adding new partnerships. You set out your last earnings call, you're going to add a call it 100 this year. So...

We did 59% last year.

Can you talk about the capacity of the foundries as you guys need to build out more, increased automation, what's going on?

Yes. So we track -- a couple of things about this, right? So -- so our -- like one of the biggest metrics, this is why I talked about this for a year, Steve, is adding new programs onto the platform because they add value to Ginkgo in 2 ways. One, they are the license for us to build bigger facilities. [indiscernible] if you don't think you're going to sell a lot more cars, you're pretty scared to build the bigger factory. You're going to sell a lot more chips, you're pretty scared to build a bigger factory with better economics. So signing up customers is our license to build bigger, more efficient factories, okay? The second reason we like signing up new deals is the IP reason because where H1 is a new opportunity to generate some data that might be what helps us get the next deal. Now we have to be able to deliver on those. So we're actually in an interesting spot. I would say the we acquired last year, acquired a company called Zymergen out in the Bay Area that brought in a bunch of new automation and software technology we had 8 acquisitions last year is a lot. And so we did kind of like an expansionary move in response to signing up a lot of new deals. This year, what I'm trying to do is keep, basically, like our expansionary spending a lid on it and really focus on driving efficiency on our platform. In other words, with the assets we have, what can we do to improve our capacity output through integrating that new automation from Zymergen driving bigger scale, improving the software, all those things. And I like this because it means that on the end of this year, I think we'll have a more like efficient platform to clone out as we expand more in the future. So that's sort of like what this year is about when it comes to it -- it comes to like our scaling. Historically, we've just kept adding. This year, I'm trying to drive a little more efficiency because I think we're like a little -- we added a lot, and I think we can actually make it more efficient this year.

Yes. Okay. Yes. All right. But...

Yes, we're signing up -- what we basically do is sign up programs, and that adds more and more demand on top of the platform and the platform team has to keep up. Yes.

Yes. Yes. And maybe let's talk about that. How are you guys able to -- if you're doing 100 a year that's going to be on every 3 or 4 days, right? How do you guys pay for all the contracts. Yes.

Yes, it's a very, very good question. A couple of things. We have -- we have like a [ dedic ], so one of the falls on our Board is Shyam Sankar from Palantir. He taught me how -- why Palantir has like a deal team function, which basically their job is to like make deal making a lot more efficient. So we created that 3 years ago now. So we have a team that's really just dedicated to making deals close after the sales team has gotten the customer convinced that they should do something with us, how do you make the actual process of close more efficient. So that's one thing. Like be very good at enterprise sales and be very good at deal making. The other thing you can do is try to make the deals more standard, like get a customer to not want as much customization in the deal, right? So we have done something along these lines. We announced Ginkgo Enzyme Services, which is basically a more standardized offering to the market, specifically around 1 type of cell engineering, improving an enzyme. And that's a much more cookie-cutter deal. It's kind of more taken or leave it. It's an experiment on our side and in making deals close faster by trying to shape the customer demand into like a square peg. But the reality is most of the market, because in, they have infinite flexibility with their internal R&D team, they want custom stock. So mostly, it's getting a fisher that closing cost of deals. And I think we're best in class at that, but you can get -- we can definitely get better. And most people look at dealmaking like one-off stuff. Like I'm doing whatever. I'm buying an asset or something from a pharma company. It's not meant to be like a machine or you're a CRO, and it's not much of an argument because you're giving away the form on every deal, right? You're like, "Oh, you keep the IP. You blah, blah, right?" Like it's -- there's not much to the deal because the CRO is not retaining much of the value. We're in between, right? Like we're keeping value, so it's more of a negotiation, but I can't have each one be like buying a drug asset.

Yes. What about like your existing partners when they expand the scope of their partnership? Is that...

That's, obviously, fast dramatically. Yes, Yes, it's much easier to do inside sales, for sure. The challenge is like you need to build a relationship with a customer before you get a lot more of those. That's point number one. But we have done that like somebodies like Givaudan, the fragrance industry went from a pilot project 4 years ago to now a big relationship. We do a lot of programs with them. We're already seeing this with some of our -- like in biopharma, it's pretty cool. We haven't talked about this, but some cell engineering is around discovery, some cell engineering is around manufacturing. So like our program with Merck, for example, is a develop enzymes to replace certain chemical steps in their API process, $144 million program. That's a manufacturing program. And so -- what's interesting is we can get into a company for either manufacturing reasons or discovery reasons and then cross-sell, either cross-sell from research to manufacturing or like if we're in 1 modality and research go to a different modality, right? We can play anywhere where there -- so I do like inside sales in the long run, but like in biopharma, particularly Steven, today, most of our new deals are new logos, right, because we just -- we're just new in that industry, right? And so our deals that we have with existing players, we don't have as much track record yet to greatly expand and it's valuable for me to get someone new, right? Like that's -- and so like we're talking to companies and everyone heard of us in biopharma, right? Like that's not the case in the chemicals industry, like everybody knows us right? But in biopharma, like almost, I bet, 20% of the companies have even talked to us. So like really -- it's early in the game.

Okay. Yes. Maybe just a follow-up on that. Is there any particular area you're receiving some more interest in therapeutics? Is it gene therapy, vaccines, RNA. What...

It's mix bag. I would say like cell and gene therapy is probably, but like we don't do a lot of antibodies. Today, there's a lot of people that do antibodies. Cell and gene is good the nucleic acid drugs like RNA stuff. We have the interesting things mature the circular RNA and stuff like that. Yes.

[indiscernible]. Got it. So then a question from the audience about Bioworks7.

Yes, sure.

So it looks like cell and gene therapy is a big push. Was it -- did you build Bioworks7 for the interest you're seeing? Or did you kind of anticipate the interest?

We kind of -- yes, so we made a bet in mammalian. So just again, maybe I'll take one step back. So I said at the beginning, right, like making a biotech drug, our argument is, it is the same regardless of exactly what modality you're in. It is finding the right piece of DNA in the right cell in the right place equals the drug. What is Roundup Ready corn? A right piece of DNA in the right cell in a greenhouse in St. Louis in 1988. Like it is the same -- like being able to design biology, we believe is completely market agnostic, okay? So let's take what we did 4 years ago when we moved and we wanted to say, "Hey, we'd only been working in bacterial cells and fungal cells. We haven't been working in mammalian cells yet. How difficult is it again going to move into mammalian? Well, culturing infrastructure is very different. The way you grow mammalian cells is very different than how you grow bacterial cells. So we had to build out totally different types of containment and things like that around certain parts of the automation and all that, okay, fine. What is not different? DNA synthesis, right? So we acquired a company called Gen9. We acquired a company called Gen9 now, 6 years ago as a competitor twist that they do in-house DNA synthesis for us. The DNA does not care if it's ending up in a plant cell, a mammalian cell or a bacterial cell. Same. What about our -- all of our metabolomics okay? We pop cells open. We run them through mass specs at ultra-high throughput. We're really, really good at this. We're doing acoustics in the mass back, all kinds of fancy stuff. Does that care? Not really. Not really. That infrastructure is extremely reusable for mammalian. So I'd say about 60%, 70% of our foundry was already useful for Mammalian. And what you're hearing with Bioworks7 is me bringing online a lot of the culturing, a lot of these things that are -- the parts that were specific to Mammalian, we needed more of that. We had done a few small work cells, like work cells just like -- again, if you come visit, you'll see it, but it's like a robotic arm in the middle of a bunch of equipment and rails and all this stuff. We've done some of those. We haven't had a dedicated pure-play mammalian facility, and that's what the new facility is.

Yes. Okay.

But in general, I wouldn't think of it like that's where we do everything associated with [indiscernible].It's just not, right? Like all of the DNA synthesis infrastructure is Mammalian and also right? Like a big fraction of what we're building now is going into Mammalian just because more of our customers are Mammalian. And nothing changed over there. All that changed was the orders going in from the scientific teams now we're asking for pieces of DNA that were designed for mammalian cells, but that infrastructure was immediately useful because it is common purpose. Does that make sense?

Yes, yes, that makes sense, yes. How big is Bioworks7? I haven't seen it yet, but I'm assuming there's kind a lot of tissue [indiscernible] and incubators.

It's mostly work cells. So like it's -- again, like because Ginkgo is not like what your mental model is, as a person with a -- like pet in their hand in a hood, but that's not how we can work at Ginkgo, right? So what you'll see is big boxes with HEPA filters and robotic arms and lots of equipment inside them. And so -- but anyway, yes, you should come check it out. But we have about -- like the whole facility down in Q4 is about 250,000 square feet.

250,000 total? Okay.

Yes.

Okay, cool. Yes. So maybe in the last 5 minutes, let's maybe pivot over to the other aspect of your business, which is kind of misunderstood [ in Bio] security. So you signed all these partnerships with all these international groups, including Ministry of Saudi Arabia. Can you spend some time on like how you're going to monetize that and what this -- what's going to be the end goal?

Yes, let's all frame it for folks. So how do I get into all this stuff. So my background to the PhD and bioengineering back at MIT, I graduated in 2008, we started the company out of grad school. And we have this interesting experience back at MIT that when we got there around 2002, where this field of synthetic biology, which is really what we're on about here at Ginkgo was getting formed on the academic side. And there's a group of academics come together, Tom Knight, who's our co-founder, Drew Andy, Pam Silver Harvard, George Church, who are sort of either engineering minded biologists or biology minded engineers. And they were kind of hanging out and saying like, maybe some of the principles of engineering and computer science will be applicable to DNA code and sort of like trying different things and seeing what were stupid ideas and what were good ideas. And one of the seeds that got planted very early was if we're successful, if we make it as easy to engineer and program a cell as it is to program a computer, that seems scary, okay, right? Biology is powerful stuff. Infectious disease is scary stuff. If we have the distribution of the tools of designing biology like we've distributed the tools of programming computers to everybody in this room, if they so desire, how do you do it safely and responsibly. And that was baked in very early on. We had all kinds of people come by, FBI, WMD all this stuff, right? And so I just want to mention that philosophical point because then along comes COVID, and we're like, "Oh my gosh, this is the moment to build our biosecurity infrastructure. Think biosecurity like cybersecurity. In other words, alongside the tools of program computers, we built cybosecurity to do it safely alongside the tools of programming sells, we're going to build biosecurity. And so Ginkgo got involved, okay? And we didn't do any of like -- we didn't do diagnostics. We helped out on some companies working on vaccines and therapeutics, but we don't do that ourselves. And -- but what we did do was things that we felt were not existing markets like surveillance. So for example, in 15 states around the country, we did thousands of schools like weekly school testing, okay? Because that's not actually -- without getting into the FDA issues. But like it is a -- the point of it is not like your sick and you're going to get care seeking behavior at the hospital and get a diagnostic test, it's just Tuesday, and you get tested on Tuesdays to try to reduce outbreaks in your school during Omicron or whatever. And so that was our first foray. And then that expanded it in a program with the CDC into similar work at airports where what we're doing is we're collecting wastewater from planes and anonymous samples from passengers coming in from different countries pacified by the CDC and looking for variance. So -- and we sequence all the positive cases. So BA2, BA3, first cases in the U.S., founded in airports program. okay? And we're now doing this, as we mentioned in a number of other countries, in Qatar, countries in Africa and so on. And the idea is eventually to have an equivalent of like radar or satellite monitoring for the weather but for the evolution of infectious disease. And we see this as not a thing that goes away with COVID or whatever. But a permanent set of infrastructure that if we're going to be grown up about handling infectious disease in a future where it's easier to engineer biology, you just got to have. And we are seeing support from governments to do that.

So they are deploying money, the different governments?

Yes.

Okay.

And so yes, so we're excited about this. It's very early, but it's going fine. And so yes, you'll see us do more of that this year.

Okay. Got it. And final question. Given your cash balance, should we expect more M&A?

Yes. So we added last year with $1.3 billion in the bank. That's another thing I think that makes us a bit unique in this current market. So you won't see us do like big -- we did a couple of big deals last year, both acquiring R&D asset from Bayer in agriculture and acquiring Zymergen. I think you'll see us some smaller things. But that -- yes, that's sort of our plan there. But mostly, it's use that cash pile to give us time to get to royalties and milestones on the deals we have.

Okay. All right. Great. Well, we're out of time. You went through a lot. I really appreciate it, Jason.

Steven, nice to talk. I appreciate the time. Thanks, everybody. Thanks.

So good afternoon. Steven Mah here on the Tools & Diagnostics team here at TD Cowen. Our next panel today is focused on synthetic-biology-enabled drug discovery and AI-driven drug discovery. So similar to other transformational technologies such as the Internet, we believe synthetic biology and AI may become one of the most important technologies of our time with the promise to address many of the global issues challenging us today, sustainability, high cost, affordability. So with technology advancements, such as gene editing, new R&D tools capable of generating huge amounts of biologically relevant data in single cells or in tissues, continuing lowering of R&D, input costs, the lowering of computing costs and increasing computing power and the emergence of machine learning and generative AI. So we believe we're at the precipice of a new revolution, innovation and drug discovery, which can deliver better and safer drugs at a faster and a much lower cost. So today, our panelists are Sean McClain, AbSci CEO and Founder; Stephen Dilly, CEO of Codexis; Jason Kelly, Founder and CEO of Ginkgo Bioworks; and Daphne Koller, CEO and Founder of Insitro. So welcome. So before we begin, I want to kind of level set the audience on the business models here. So there's 2 classes of business models in the space here, which are not mutually exclusive. So some companies employ a hybrid model with elements of both. So the first business model is lab to market where this is more of a traditional biotech model where people use their platform to discover drugs and enter them into clinical trials. The second one is what we've coined R&D sharing economy business model, which involves executing programs for partners in exchange for upfront licensing fees and then more importantly, getting downstream value in the form of technical milestones, clinical milestones and royalties on commercial sales. So on the panel, let's keep it interactive as much as possible. I'll be checking my phone for any audience questions as well. If you want to e-mail me at [email protected]. So with that, let's begin. Given the framework I just kind of laid out on business models, can each of you spend a few minutes giving an overview of your company and business model and also focus on what therapeutic areas and what type of modality for these small molecule antibodies, whatnot that you're focused on? So yes, maybe we'll start with Sean, then Stephen, then Jason, then finish up with Daphne.

Yes, absolutely. So AbSci is a generative-AI drug creation company, really going from this paradigm of drug discovery where you're finding the needle in the haystack to drug creation where you're actually creating the needle. And in our case, it's the actual biologic, being able to leverage our wet lab technologies and capabilities that generate enough data to actually start to train these generative AI models. And essentially, what we're doing here at AbSci is doing what DALL-E has done, being able to go from text to image. But instead, what we're doing is going from target to antibody, being able to specifically identify the antibodies with the attributes you want the first go around, being able to go after some of these undruggable targets, GPCR ion channels, that have been difficult to drug. And now with technologies like generative AI, you can actually start designing these in silico and really unlocking new areas of biology to go into and dramatically shortening the time it takes to get into the clinic. And our business model is a hybrid business model where we both partner, but we're also developing our new pipeline, especially with Andreas Busch coming in who was the former Head of R&D. With him coming on board, we've decided to start building out some of our own pipeline. And we're actually doing this to really create a flywheel effect for partnerships. When -- we've demonstrated in the lab that we can design antibodies de novo on a computer. But what we have yet to show is a clinical proof-of-concept. And you're always going to have early adopters of technology. But what you ultimately want to do is get to that clinical proof-of-concept as quickly as possible to then have adoption of more large pharma coming in and creating that flywheel effect. And so we're -- and obviously, there's intrinsic value of the pipeline itself.

Fantastic. Thanks. So I'm Stephen Dilly, CEO of Codexis. I'm a very visioned drug developer, I think, doing for 35 years or more, have failed in just about every area it's possible to fail and succeeded in a few, learned a lot of hard lessons. And so in August last year, I came in as CEO of Codexis and found myself like a kid in a candy store with this wonderful technology where we can do directed evolution of enzymes using AI and machine learning, but also using real intelligence as well in terms of starting points from where we're going and saying, okay, what's that for? And what we're focused on is right at the other end of the spectrum is making better [ mass traps. ] So we're looking at existing moieties -- existing clinical moieties and saying, we can tweak that, we can do a little bit better because it's derisked, there's a real route to market. And so we're starting walking and then we will run later. But it very much is saying, there's a real problem in the clinic that we can address with our technology right now, let's work on that, let's validate ourselves, and then let's move on to the more ambitious stuff. And right now, we have 2 very big partnerships. One of them is with Nestlé Health Science where we're working with the metabolic group around things like pancreatic enzyme replacement therapy and inborn errors of metabolism. Another one of our big collaborations is with Takeda, where Takeda has said, in gene therapy, we want to be better rather than first, and we're using our technology to engineer transgenes to go on their platform. But again, people know what the problem is with the lead agents in fabric. We know what the parameters are, we're trying to address, so it's very directed.

Great. Jason?

Sure. I'm Jason Kelly, Co-Founder and CEO at Ginkgo Bioworks. I guess, Stephen, [ your part of lance would ] be the R&D sharing economy model. Yes, so we don't develop our own drug products. We don't have a pipeline here. We're maybe unique in that. We also work outside of biopharma. So we also work in sort of industrials and agriculture. And part of the theory of Ginkgo is you simply can't vertically integrate into all those different markets, right? I cannot be developing new ag biologicals and the new cancer therapeutic, add a new fragrance, even though I work with Givaudan in fragrances and Biogen in AAVs and Bayer in agriculture, right? It works -- I am more scalable because my customers are ultimately responsible for the -- to bring products to market. And what they're outsourcing to me is R&D work and they're doing it because they want access to the unique assets at Ginkgo, and there's really 2 of them. One is very, very large-scale automation. So if you want to visit [email protected]. We are about 10 minutes away from here down the seaport. About 250,000 square foot, highly automated lab. It is pretty cool to see. Really, you get a sense of the scale and what you can do with it. So data per dollar is better at Ginkgo than in your own in-house lab at -- again, at Biogen or Merck or Novo Nordisk, all customers of ours. And then the second asset is the existing data I have from all the work I've done previously that I used to inform the design of an engineered cell in these different markets. And so we're learning every time we do a project for a customer, that trains our models, AI models, but also just design theory, right, and sometimes also physical assets like genetic assets, whether it's in capsids for AAVs or circular RNA. We have genetic code assets for previous projects we've done that we can bring to Bayer for new customers in those markets. And so I can talk about that model. But that's the business. So yes, R&D sharing economy, yes.

Thanks, Jason.

I'm Daphne Koller, I'm the Founder and CEO of Insitro. And I guess we're an outlier in this panel as well because we're fundamentally a biology discovery company. So what we do differently from most of the AI and drug discovery companies out there as we unravel the complexities of human biology in order to uncover novel targets and novel patient segments in which those targets are actionable. We also generate therapeutic matter against those targets, but the core of what we do is really identifying novel hypothesis that can drive impact to patients. We do that by combining 2 forms of data, some of which we generate in-house and some of which we acquire via partnerships with others. The first is cellular data. We also have a highly automated lab that generates biological data at scale, specifically the relationship between genetic interventions and their phenotypic consequences in disease-relevant cellular systems and clinical data, which is the other half of the equation for us is ones we acquire via partnerships with others so that we can interrogate the effects of genetic interventions, experiments of nature in this case of human pathophysiology, and we use machine learning in order to bring them together and cover novel hypothesis. Most of the work that we do is towards our own pipeline. We have had a couple of transformative partnerships with biopharma that enabled us to build our platform in a partner [ and ] asset format, but the goal is to bring our own therapeutics to market in a way that gives us things that have a higher probability of success in patients.

Okay. Great. Yes. Thanks for that. So that's a good segue to asking about R&D productivity gains driven by AI machine learning. You mentioned a bunch of different tools or databases there. So maybe just discuss what's enabling these productivity gains? Is it the algorithm? Is it the data? Is it your functional data that you feed into the machine learning algorithms? Just give us a better sense?

First, I think, it's useful to disentangle productivity gains into 3 facets. There is better, there is faster and there's cheaper. And each of those are important. Better means you put stuff into the clinic that has a higher probability of success, which, to my mind, is probably the most important because the big costs in our industry are all about push stuff into clinic. You make a very -- you do a very expensive trial and doesn't work. There is faster, which is how do you progress faster in your preclinical and then subsequently potentially in your clinical stage programs. And then there's cheaper, can you lower the price? The fundamental focus of our work is on better, which is one to increase the probability that when we put something into a human, it will actually make that human better. Part of what we also benefit from that is if -- given that better also involves identifying the right patient population, you also have the opportunity to go faster because if you can narrow your clinical trial on the right patient population, you get larger effect sizes, you can reduce the size of your trial and hence, the cost of your trial as well, but it all emanates from having the right clinical hypothesis to be [ in way. ]

Great. [indiscernible] let's keep it interactive whoever wants to go, just hop in.

Got it. I definitely agree with you in the 3 different areas. And at outside, we're really focused on more on the biological design. How can we design better antibodies that are going to increase that success in the clinic and get them into the clinic much faster? And really what's driving partnerships like Merck that we had closed last year and others that we have in the pipeline is really the fact that we can start designing antibodies or biologics that existing technologies can't design for. And let's say, for example, a GPCR or an ion channel of a particular epitope that existing technologies like immunization won't generate for you. If you're able to start unlocking this, you're able to start unlocking new biology. And this is what pharma cares about and this is what ultimately is going to get better drugs into the clinic faster. And then additionally, if you're able to design everything ultimately fully in silico on a computer, you're going to be able to get into the clinic much, much faster. The -- we just started work on our own pipeline, as I had mentioned. And we are anticipating to have an IND at the end of next year, and so that will be roughly 18 to 24 months from start to finish on an IND. And that's really compared to 5.5 years that it takes pharma. And so that additionally gives you an extended patent life. And I think these were the things that our pharma partners get really excited about. But at the end of the day is what's going to ultimately increase success rates and decrease the amount of time it takes to get new therapies into the clinic.

So I'm going to take it down a level in terms of complexity and say that, to me, success is about commercial traction. And so you actually have to start right at the end of the process and say, can we do something useful and will someone pay for it? So actually, one of the most important inputs is understanding the market that you're trying to address and what the pain points are and how you can actually solve them. And I think in all of these sort of platform companies, we have a wealth of opportunities, and it's about selecting the few that we're going to take forward aggressively because we can do cool science to look blue in the face. And unless there's a regulatory endpoint that we can nail, unless there's a reimbursement path, unless there's a clinical path, we're not going to get very far. And so I think there's enormous promise in what you're saying, but sometimes it is the prosaic stuff that gets in the way is -- because we've got to drag the agency with us into the right century, right, in terms of the endpoints we're going against, in terms of you can get confident around in silico development. But there's a regulator out there that has to do as well before they're going to let you expose patients. And now one way we address that is going for a very high unmet need where the risk is worth it. But what do you think?

Yes. I mean I -- look, I haven't had 1 pharma partner tell us that they're going to have regulatory concerns. If we can do something that they can't do and other technologies can't do and that achieves the biology that they're looking for, that's what they're going to take the bet on because they need that next blockbuster drug, and they can use this. It doesn't matter whether you can do it in silico or with Synbio platform. At the end of the day, what you're delivering is something highly differentiated. And that's what matters. It doesn't matter whether it was done with AI or in the wet lab, it's how do you deliver something that's highly differentiated. And that's what ultimately drives pharma partnerships and ultimately leads to the success in the clinic and makes you highly differentiated. .

I think ultimately, the regulator doesn't really care if it came off the computer, out of the lab or someone hallucinated it in the middle of the night, you have a preclinical package that you submit to the regulator and either it satisfies the requirements or it doesn't, how it was discovered, I think, is not something the regulator particularly cares about. I think the pharma partners and we care about how it was discovered because we may have discovered it better or faster, but ultimately, what matters is there's a preclinical package that the regulator looks at, which is pretty standard.

So you're implying that what I was hearing pure in silico development is a long way away there. .

No, I think what I'm referring to is that the way you discover the antibody is in silico, you give it a target, you get an antibody, but you still have all the GLP talks that has to be done. We're not saying that that's - and the efficacy studies and all that. That work -- that's the 18 to 24 months I'm referring to is -- but everything past that is all done in silico, but we ultimately still have to do GLP talks and all of our efficacy studies.

The only thing I'd add from the R&D sharing economy over here at the end because people ask you all time like, which programs you be winning programs that can go how do you think about which -- whatever drug targets depend on. And I think one of the features of us not having a product pipeline is we don't have to care, right? What we're really striving for is ubiquity, right? Our view is that activity that today is distributed widely throughout the industry. In other words, every company has their own R&D lab running at low throughput with scientists in white lab coats, holding [ pipe bats ] working by hand miserably at the bench, should be centralized at 1 big automated facility, and we should take a very small piece of many pies, and I don't have to pick the pie. And so that is the one difference if you're adopting just a pure-play research services model.

All right. Thanks. So you talked about fully in silico, I'm going to get to that later. But I don't want to jump around too much, but I want to stay a little bit higher level just for time being, but pay no attention to that. So like every company will say they have the best database, say they have the best algorithms. How can investors differentiate the companies that are saying that, right? So how do they sift through the -- find the wheat from the chaff?

I have my opinion...

After that, I'll try it -- last time -- first and last -- So we -- my view is that a lot of the algorithms are pretty commoditized at this point. Lot of people have a big differentiation there. It comes down and do you have proprietary data assets. And so you have to get those somehow, right? Like they can be through partnerships, some people that have certain things that's particularly true on the clinical side. But also, I do think your ability to generate data that's proprietary ends up being a big differentiator here, in my view. I think that's lost on people that are further from the technology, but I think that's the game.

Okay. So I'm going to largely agree with Jason and say that having proprietary data of high-quality data that fits best for machine learning is actually really important. And there's a lot of companies that think that if you just kind of make a big pile of data from here and there and everywhere, then somehow magically machine learning will be able to make sense of it. And what we've seen is that machine learning applied to a pile of, shall we say, heterogeneous garbage produces garbage because it is the same ultrasensitive lens that machine learning can use these days and cover several signal, equally well allows it to home in on several artifacts and that those are rife in those heterogeneous data sets. Having said that, I'm going to slightly disagree with Jason in machine learning becoming commoditized in saying that while certainly machine learning core capabilities are commodified, how exactly you take the data and put it together and ask the right questions using the right algorithms. That is not commoditized. And so maybe coming back to your original question, one of the other big differentiators that an investor can look forward in deciding what is a good or not good investment thesis is the quality of the team and do they really bring together the right sensibility on the machine learning side and the qualifications of really bringing machine learning at the frontier together with understanding biological questions in depth, so that you're applying the right questions, the right answers. And we've seen that recently in some of the work that we've done in applying to highly investigated public data sets that when you ask the questions in a different way, sometimes incredibly novel insights emerge from that. So I would say, yes, high-quality data is absolutely important, but also is having the right people who ask the right questions in the right way.

Okay.

And I would say another really important aspect is being able to iterate on your hyper parameters, the model architectures. And that means being able to generate data to solve the questions you're asking. You got to be able to be go in the wet lab, get the data, train your model and then go and validate that and then be able to start that process all over again. And so being able to have very rapid cycle times where you can go from data in the wet lab, training your models and validating is super important. And I think that's one of the reasons why we've made so much progress because we can do that in a 6-week time period. And not only that, we're actually wet lab validating the model designs. We can validate 2.8 million AI design antibodies in a given week. And that allows us to inform the next model designs and architectures. And actually, we're discovering new areas within machine learning to go in it. Yes, it's based off of what's been out there, but there are novel insights that our AI teams are developing all the time. And so I do think it's a combination of both, but it's not just quantity to Daphne's point. It's about being able to tailor it and also to the insights to the questions you're asking.

Right. So Codexis has been using machine learning and AI on enzyme evolution since 2007. And what we've learned is that you need to constrain the challenge. You also, to echo Daphne's words, you need the right people in place because there's a judgment call in there as well about what you're seeing and how to respond to it. It's also the quality and the size of -- it's the quality over the size of the data set. So in the constrained problem set of enzyme evolution if you've been doing it for 20 years, then you tend to have more informed starting points and more informed guesses about the direction you're taking, and it's really that kind of thing rather than -- I said how old I am before, but I've been through so many bloody revolutions, which are going to change the entire landscape that what you learn is, over time, yes, they have utility, proteomics, genomics [indiscernible] I can go on forever, right? They'll have utility in well-defined spaces for specific problems, and then we will move on, and we'll get excited about something else. And I'm too old to believe in all-encompassing sort of revolutions for everything.

But I will say though, in order to get the best talent, you have to have the data and that data engine -- I mean we've recruited top AI talent from Tesla, from meta, from open AI. And they come here because they know that they have the ability to solve this big audacious problem because we can -- they can specifically ask the wet lab to generate the data that they want in order to train their models and validate it and rapidly iterate, they basically are at a tech company but in biology, and that's what's attracting the best talent because there's so much that can be done here, assuming you have scalable wet lab technologies that get you that data.

Yes.

So can I just slightly disagree with Stephen in the following way which is to say that while it's true that there have been point technologies that have made an impact in sort of certain slices, if you will, to drug discovery and development company, and I've heard people make a similar point as the one that you just made, which is that machine learning, AI will be of that kind. My analogy for what machine learning will be is a little bit different because machine learning will be like computers. They will make a difference in every single part of the drug discovery development company. Now is that going to be [ silver board ] solution for that? No, but it will not be a point solution that is relevant just here or just there, I am already seeing machine learning making impact on every single stage of the pipeline of drug discovery and development. The extent of the impact depends on the quality of the data that is available in that slice and the quality of the people that are applying those data, but it's not going to be a point solution.

Right. I actually completely agree with you. But it's -- in each of those stages, it's going to have a defined impact and those are going to join up and it's going to make a fundamental difference.

Exactly.

I mean it's like looking at a computer and saying, well, that's going to solve my problems for it.

That's exactly right.

Yes.

But it is going to make a difference everywhere, not just in a single slice. I think that is what makes a difference for a lot of the other technologies that you mentioned because it's going to be ubiquitous in the same way the computers are ubiquitous.

It's the computerization of medicine.

We'll go down here at the R&D sharing economy. I would tell -- we -- one of the cool things is you get earlier signal as to where the applicability of things are because you need to get like a customer to sign up in order for us to sell our services, right? And so if it's not element in an area, we find out quickly because no one will do a deal with us, right? So if you look at, I guess, now 96 active customer programs on the platform, 1/3 of them are in biopharma, 1/3 of them are in ag and a 1/3 are in like industrials and consumer. So I think for some of these technologies, our case like a high-throughput robotics and some of these data assets, I think you are seeing true breadth. I am sympathetic to the 5.5 years long history, I'm like this is going to change all of how pharmaceuticals is done. But some of the lower-level technology. I think it don't touch everything. It's a question of how can you then commercialize them, right? And if you're pursuing a drug like the idea that you'll make a drug in everything as a result, this is probably not right. But could it somehow be applied broadly, I think the answer is feeling like yes for some of these things.

Well, no, that's a great segue way then to a question I got from the audience. Specifically for Daphne, but I think everybody -- I mean Jason, you can answer this as well. But the question was there's a difference between discovering biology with your databases and proving that there's a biological effect. And then what strategies you guys employ to show proof-of-concept before moving a molecule going downstream? Is that...

So I think that's a great question, and I think maybe we'll come to some of the questions that you were going to ask in a few minutes about purely in silico discovery. We don't do purely in silico discovery. We generate data in the lab, we collect data from human patients. And then we use machine learning to interrogate those data to come up with plausible ways -- possible mechanisms, possible targets in patient populations. And yes, ultimately, these have to be validated and tested. And one of the questions that every biopharma asks when interrogating a new target is how do I gain conviction around that, and that is always a process. And so there's ways that one can use experimental data to do that, and we have experimental systems that we built in-house that we feel give us considerable conviction around the viability of the target. Some of those are earlier stage more broad-based. Others are very targeted and have a much greater level, maybe smaller scale that have a higher level of translatability. So for example, some of our neuroscience programs, we have model systems that we've built that measure electrical activity in neurons, and we see the mirroring the electrical activity in epileptic patients. And so when you have a drug that actually modulates that activity and reverse it back to something that looks more like normal, you have a greater conviction that, that is a translatable assay. The other form of forces, you can also use human data in order to gain conviction. We've all heard the statistics. The drugs that have -- those targets have support in human genetics are twice as likely to succeed in the clinic. So our ability to interrogate human clinical data with an eye towards genetically validated targets with machine learning to really interpret the phenotypes that we're seeing is another way that we have that will give us conviction around this. And then ultimately, the truth is that the only time that you have true conviction that your drug is going to work is when you put it in the patient and you've seen that it actually responds. And so that's called the clinical trial.

One question is for -- this is for anybody. But do you guys use some of your AI machine learning to -- I mean we've been talking about discovery and target, right? But you ever use your databases and systems to fail something faster because, ultimately, that will save money as well because you're not pursuing things down a rabbit hole. Is that a commonly used? Or is it more just finding targets and then pushing them through the -- finding the...

Well, we are finding targets. I'm not sure that that's true. I mean we find targets. That's...

No, no, I'm just curious...

I'm [indiscernible] on the panel and that...

We use it to identify blind alleys in what we're trying to do to say, look, we can't go in that direction. But this is a wonderful conversation because we're almost behind key matter to what Insitro is trying to do in terms of...

Don't say that...

We have a table that's going to be asked, so...

But what -- okay, we're wonderful team -- So because what we're doing is taking drugs that work and making them work a bit better. And so there where we're using the AI and machine learning is saying, can we take this orally administered enzyme, tweak it and evolve it and make it pH independent? So it can -- I'm talking about pancreas enzyme replacement therapy, lipase, where the problem with orally administered lipase is they get chewed up in the stomach, and they don't have the potency we want. So we engineer intrinsic potency. We engineer pH independence and survivability. Those are parameters where we've actually got an assay platform that we can iterate fast against. And where we're using the machine learning, the AI is informing the starting point, the directions we can go. So it's not just blind evolution, it's directed evolution. But then we're coming out with something that we can test very rapidly because one of the other lessons that I've learned in drug development is a developable drug has an early readout on whether it has a way forward or not. The worst of all is the drug that reads out at the end of Phase III, and we're all terribly disappointed. So that's the way it's really selecting things that can be early kills on clinical data.

And so for sure, we do that for our own targets. We run the gamut for them through a variety of different ways and we try and kill them off as quickly as possible. So we ended up with ones that we think are likely to succeed. If your question is, do we try and look for things that are in biopharma pipelines and go tell some of their program is not going to work? No, we do not do that.

Yes. Maybe one thing I would add. I think there's a general challenge across the industry that there is a lot of learning that ends up sort of stove pipes across different companies in the sector that could save somebody else from making a mistake. And so one of the things that I'm kind of hopeful happens with some of these platform business models as they allow for an aggregation of IP and learning across programs at different companies, even across competing companies in a way that ultimately is healthier for the entire industry. And so you ask to come with a little bit of a cost like someone else is finding out something but net you're getting more drugs to market, you're getting more ag products to market, whatever industry it is. I think in bio, a lot of the IP regime and mental model around data has been counterproductive from a collective industry standpoint. And some of these platforms will hopefully be powerful enough to force a change in that IP regime. That's certainly what we're hoping to do at Ginkgo.

Great. Last high-level question, then we'll get into some more detailed questions. But let's talk about public databases. I know that those will come up. I'm talking about like there's genomic databases, there's also these kind of open stores, foundational AI models, AlphaFold 1, AlphaFold 2, OpenFold. These big tech companies have invested a ton of money in these sort of open-source models that are pretty much available to anyone. So my question is how are you guys capturing this value? And is there a risk of these open-source model sort of commoditizing the industry?

Look, I think the open-source models are great. I mean I think what AlphaFold did for this industry is phenomenal. I think what it showed was the power of AI. But if we're going to take it to the next level, you need data and lots of it. Let's just take a look at OpenAI and ChatGPT. Right now, they're investing a lot of money to do reinforcement learning on improving the accuracy of their models. They're basically hiring people to generate data to ultimately increase the accuracy. And if you look at like what we're doing here at outside, I fundamentally believe that you can go fully in silico design everything on a computer. And we've demonstrated that we are -- it is possible. But right now, where we're at is that you can use AI, throwing the structure of your target, you get hits that come out of your model and maybe 10% of the time when you go and test it in the lab, they're correct. But in order to go fully in silico, you need to generate a lot of data that ultimately increases the accuracy. So you go from being 10% accurate to 50% to 99%. And you can't get that with the public data sources. You have to specifically tailor data to ultimately get fully in silico. And that's the journey that we're on here at outside. But you're not going to get there with open-source data that currently exists. I think that you enable yourself to make huge advancements. But to get all the way there, you need a mass amount of data to train your models just like ChatGPT and what OpenAI has done.

But are you guys taking this open-source data and then using it?

So for our purposes, the data that is most useful and relevant human clinical data and one of our favorite data sets that we've made extensive use of is the U.K. BioBank because it is one of the very few very high quality, very highly harmonized human clinical data sets out there that is readily available. The amount of value that we've been able to extract from that is quite remarkable. And what we are now doing is working to identify other similar resources, many of which would be proprietary so we can gain via partnerships. Our partnership with Genomics England on Oncology is another incredible example of a wonderful data set that we have differentiated access to because of the nature of the partnership and there's others that are in progress. So those are the data sets that are helpful to us because, as I mentioned earlier, our focus is on identifying novel clinical hypothesis and those are the data that are most useful for that.

Yes. We also love the open-source data sets. They give us much better starting points when we can -- often to actually step back to your point, they often exclude things because we can look at it and say that's not -- adding the human judgment to it, that's not likely to work down that alley if we can look at the likely structure -- one of our sidelines a bit like Jason has a number of different businesses is in pharmaceutical manufacturing, where we make enzymes to catalyze specific reactions for pharmaceutical intermediates. And that's a place where this really does play in because that's about taking a specific chemical structure, gripping it in the pocket and turning it into another by known reaction like, for instance, a keto reductase or something like that. And that's a place where the existing databases can have real utility in terms of saying, how do you design that pocket to have the right starting shape so you can refine around it. So I'm not total unbeliever...

All right. All right. That makes sense. All right. So let's finally talk about fully in silico, and yes, this is more for -- more biologics. But what are the limitations of doing a purely like in silico antibody and compare and contrast to people which are using natural antibody approach, either immunization regimens and immunized animals or whatnot. Can you guys compare and contrast what...

I can take that -- look, if you look at how biologic drug discovery is done, particularly with antibodies use, [ features play ] or immunization. Let's just take immunization, for example, you go and take a humanized mouse, you inject in the target or antigen and the mouse generates antibodies for you. But you can't tell a mouse to generate an antibody that binds to the epitope that you want, the region on the target. You can't tell it to have the affinity or functionality you want nor the manufacturability or developability. And you have to go through this long iterative process where you change the sequence and then that sequence actually then changes another attribute. And so you go through this long iterative process. I mean this is one of the reasons why it takes 5.5 years to get new drug candidates into the clinic. And when I talk about folding in silico, the vision that we have here at AbSci, and it will become a reality is being able to specifically target all the attributes you want in a zero-shot manner. You take a target that you're interested in generating an antibody towards. You may not know the biology to this novel target, but you can generate an antibody that will bind to all of the epitopes on the target at various different affinities and you go into the wet lab and you find out what -- which of these gives me the biology that I want? And then ultimately, you can start training with that data to ultimately get to knowing the biology, the first go around. but this is a huge advancement and being able to do that, that's what's going to unlock new biology, being able to hit epitopes that you've never been able to hit before. And it's essentially kind of like the ChatGPT of antibody discovery. But this is the future. This is the computerization of medicine and where we're headed as an industry. And we recently came out with a preprint demonstrating that we can do exactly this. We're not fully in silico at the moment. Maybe we have 5% to 10% hit rate, but we're continuing to train with the data to ultimately get to the point where you can specify the epitope you want and the affinity. And there's -- you basically eliminate any sort of wet lab need, and you can basically take it straight into GLP talk studies from there.

Okay.

I'll add a little bit of this outside of antibodies. Yes, Ginkgo philosophically was sort of a bunch of engineers that came into biology when I started the company. And so there's like a certainly a dream of ultimately getting to -- think of it like computer-aided design tools for semiconductors, right? Like you design a whole very complicated semiconductor chip, it costs a fortune to manufacture even a prototype 1, and the thing just works, right, because the CAD tools are perfect. And so that's sort of an aspiration there. I think what's fundamentally unique about biology is that like we did not design the stuff. We inherited it from 4 billion years of evolution. And so I think what the way this game is going to play out because Ginkgo's box is basically changing the genome of a cell, right? We're going to change the genome of a cell for any purpose, we think you could outsource that to us and we'll help you find the right genome. All right. I think what will happen is pieces of the problem will start to become predictable. And then you as a product designer will say, "Great. I now know how to design that. And I would like to build a more complicated thing that I don't know how to design yet, and I'll be back to needing to try a lot of designs." And so you'll have this like, we will climb a curve where we'll start to employ what computer scientist would call abstraction. And a certain part of the problem, you'll be able to abstract away into computerized design. And the next level, maybe now you want to design a whole cell, whatever it might be, is going to once again kick back to needing high throughput in the lab. And we'll just go and go and go for decades along that train and get better and better at designing pieces of biology. I think you're seeing awesome stuff with AlphaFold, we're like, hey, we can start to get [indiscernible] hands around okay, right? Obviously, cell is made up of millions of proteins interacting, right? So that's the -- I think that's our collective journey in my view in the coming decades.

So just a word of caution is the ability of biology to come and bite you in the butt when you don't expect it. I mean thinking about some of the glorious failures have been places where we thought there was almost no risk. And these 4 billion years of evolution have built a lot of feedback loops and a lot of compensatory mechanisms. And going from the abstracted element to the impact in the complete system is still a long way away, right?

But I think that AI, ML is going to accelerate our understanding of biology in a way that we haven't ever seen. I mean even the designs that are coming out of our AI, it's crazy. I mean, we can change -- like change the CDR3 loop and an antibody, the most diverse area of an antibody to be 90% different in the sequence, but you still have the same structure. And you don't see any of the sort of stuff reported in literature. And if you look at evolution, evolution happened in a particular way. And I think what you're starting to see is AI actually helping us understand not only the biology but also looking outside of kind of the evolutionary space that, that was ultimately created. So I think it's going to help us accelerate our fundamental understanding of biology. And I'm really excited to see what that entails.

And so maybe to bring that to the world that we live in, I don't think anyone at this point is seriously thinking that you could just build a model of the human body. And so for a prediction of what target is likely or not likely to work in that. I think the closest that we've come as a community is to identify genetic drivers of disease using human genetic analysis. And I think those demonstrated to be better drug targets, but I don't think anyone thinks that you can just take one of those and just say, yes, I have full conviction around that, and we don't need to do any additional validation of that. And so I think that there is -- we're getting better and better at disentangling which of those genetic associations, if you will, are more likely to be meaningful drug targets and pursuing the ones that are higher probability of success, but there's still a validation that needs to be done on there, I am talking about.

Okay. Great. That's a great perspective. I get a question from the audience. And Stephen, this kind of goes to you talking about biology biting you in the butt. And then also to my question about failed, can you learn from uses to fail faster? So the question is, has anyone ever used like failed clinical trials as like a starting point or mine that data to instruct new clinical trials or new drug development?

Yes, I mean there's....

Aggregating that data.

There's been a bunch of papers that have tried to apply machine learning on various features of clinical trials to predict success or failure of the upcoming months. They end up using some really interesting, let's just say, nonbiological features like who is the PI on the -- and how quickly will they put to recruit and things like that. So from a hedge fund perspective, it might be a useful business model; from the perspective of understanding underlying biology, maybe not so much.

And one of the classics of that is to look in a population where overall a drug failed and try and find the subgroup where it worked. And that is the last [ refuge ] of the dam, and I can say that because I've done it, and I've done it in Phase III, and it usually comes and bites you because there's something external to the data set you weren't predicting. But I think it is -- I think it's really useful in near misses where there was clear evidence of activity, but not enough. And the thing that still haunts me 20 years later is I was part of developing IL-2 -- high-dose IL-2 for things like multiple -- for melanoma and all kinds of different areas. And I looked on it as a failure because we had an 8% response rate. But 15 years later, I was on a panel sitting next to a charming 45-year-old lady who was 15 years after high-dose IL-2 and was a survivor of end-stage melanoma and how do we find who that 8% is because you have to take them [indiscernible] but it's worth if you're going to get that, right? And so that's an area where I'd see very...

I agree, and I think that's a very important area because we have made some very, very good drugs that have not been successful because we have not identified correctly the patients that are more to benefit. So I think one of the most exciting uses of machine learning in the coming years is going to be correct identification of responders to drugs that are actually affected, but not effective on an all-comers basis. So just as an example of some colleagues of mine at Calico who works and I think at the time that Herceptin was developed, said that, had they applied Herceptin to an all-comers breast cancer population? It would have failed, we would have needed -- you could have had -- 15,000 patients and the curves were actually right on top of each other because the harm done to the patients who are not HER2-positive would have outweighed the benefits of the HER2-positive. And so the curves from a value creation perspective would have been actually not -- you would not have been able to get a permission. And I think that to me is a great example of when you need to correctly identify the patient's population of the right biomarker.

Yes. I totally agree with that. Yes, I have a theory as well that all these Alzheimer's drugs tend to fail because you're not currently diagnosing -- done because...

Alzheimer's is not 1 disease.

No, exactly.

Multimodal as well, yes.

Yes.

Okay. In the last 10 minutes, I got another question here. So most of you've been talking about biologics and protein, gene therapy. But the question is, can AI machine learning be used in other drug modalities, small molecules, antibiotics, inhibitors of antibiotic resistance? Or do you guys know of any companies in that space? I personally don't, but...

There's thousands all over...

Yes, for antibiotics?

Or antibiotics.

Small molecules...

Antibiotics is like a business model problem rather than a technical problem I would say. The issue with antibiotics, like no one wants to even think past clinical trial and they get funded. So it's a separate can of worms.

But for small molecules, there's thousands of companies...

Antibiotics, specifically inhibitors of antibiotics.

Well, I can chime in on this, not for AI, but for synthetic biology. So I think there's a lot of opportunity in mining natural biology for new antibiotics via basically large genomics databases of microbial sequences because when we went looking for antibiotics the first time and there's some companies that have tried this like [indiscernible] company that actually can go and acquire some of the assets of. But people went hunting by basically looking at what microbes grew in the lab and seeing if they could find antibiotics back in the 60s and 70s in the big pharma companies. But modern way to do that would be to look at the genomes and then go look using AI and computational tools for candidates, pathways and then go synthesize that DNA and express it and see if the molecules of interest. And we've had programs like that and Ginkgo is brilliant. I think the general challenge of antibiotics though is that no one wants to pay for it. So that's a real challenge. But I think the technologies are actually there to do it. And I think it would be smart from almost like a biosecurity footing for us to have a few extra antibiotics in the tank because we're -- you don't want a multidrug resistant. We don't want to go back to the era where bacteria took people out. It was not a fun time. So...

I think -- yes, I think there is a company, Biomet is doing it for animal health and looking at antibiotics as well. But again, different marketplaces, it's for production animals versus -- market, yes.

Sure enough.

Yes. Yes, fair enough. Yes, okay. We're rapidly running out of time. So let's talk about talent. I mean Sean, you talked about you guys are being able to recruit people from these big tech companies. But does it matter that these people don't have a life science background? Or vice versa, you have biologists that don't understand the programming? And how do these 2 disparate groups work together?

So I think it's really important to have talent that fundamentally understands both. And it also depends on what problem you're trying to solve. If you're, let's say, trying to solve kind of the antibody lead optimization, where you're needing a lot of wet lab data versus working on kind of new fundamental model designs on kind of the de novo, there's different skill sets that are ultimately needed. But ultimately, we do try to have candidates that do have backgrounds in biology. And when I -- 2 years ago when we acquired de novo to bring in our AI capabilities, what I was actually shocked by when we went out and started getting the AI talent was these large tech companies actually have research groups that are looking at protein design. And it's all the way from, obviously, you have Meta, Google, but also to Salesforce and some of these like players that you would never think. And so there's actually folks that are getting trained on both the biology as well as the AI. And to me, like that's absolutely critical having both of those skill sets. But we've been successful in recruiting. And I think the reason, again, why we're getting the best talent is because they can iterate on their model designs and architectures just like they would at a tech company. And they know that we have the ability to actually solve this problem and they want to be a part of doing something that's big.

And so I'm going to put a point on that and say it's not just about the talent you bring in. It's about how you connect them up and it's setting up the forums for information exchange because quite often the nonbiologists is super intrigued by biology and vice versa and setting up the right conversation in the right context and encouraging it, actually you can bridge a lot of those gaps.

It's that integration. I mean you have to have the wet lab and AI integration. And I took us 2 years to get to the point where we can go from billions of data points in the wet lab to training to then validating 2.8 million designs in a week, it took us time to build that, but that is so key, and it's having each party fully integrated in order to achieve it. You can't have them as like 2 separate groups.

To all those points, I think the culture in these companies, this new age company that really bring together, these different disciplines is absolutely critical because people will naturally self-segregate into isolated groups where you sort of developed the growing stuff over the wall. And so building in a culture where engagement and collaboration as part of the fabric of the company is absolutely critical. We hire to that mindset. We hire some people who don't speak the other language, but there's a common core of people who what we think of as bilingual scientists who actually speak both languages and they form a connective tissue between the people who are on one side or the other, and that creates a continuum. We're seeing flow through the company versus people naturally segregate with their own kind, which I think is a real failure mode for this kind of company.

Okay.

And I think you need a mindset where you want people to learn both at the end of the day, it was so funny, it was like 9 p.m. the other night on at our campus and one of our AI scientists was intown and I walked by our townhall area, and he's got like a textbook out and like -- what are you doing? He's like, Oh! he is like, I'm just -- I'm studying up on biochemistry on structures. I was a little weak in some of these areas. And it's like people that -- I mean he's like a world-renowned AI scientist, but like he is fundamentally curious and knows how important it is to understand like both and you have to have that kind of curiosity to understand both to be successful in this new era.

And it's the feedback loop as well. It's that your attachment doesn't end at a certain time because there's information coming from success or failure of what you've pushed over the silo, right?

Yes, absolutely.

Great. Yes, so in the last couple of minutes, final question. So synthetic biology, AI, machine learning, is it become buzzword, to all these publications out there. You've got ChatGPT, DALL-E, all hitting the mainstream now. Is this hurting or helping your cause? I mean there's some people say this is the next big bubble.

It is not the next big bubble. And sorry...

You go ahead and complete.

Yes. No, no, no. I will say it's not the next big bubble. And I will say like ChatGPT, DALL-E, like what open AI has done for us is actually help investors understand like the impact of this. If you can tell an investor and even my parents like, I think, finally, fundamentally understand like what I do. It's like, hey mom and dad, it's like -- instead of going text to image, I'm going drug target to antibody, you start to understand the impact of this. And I think that, obviously, there's still a lot of work left to be done here, but I think that we've fundamentally shown that AI is here to say. Generative AI is not just a buzzword, and it's all going to -- look, I mean, we all have gone on and use ChatGPT, whether it's to ride a wrap on AI drug discovery or using DALL-E, we've all used and seen like the impacts of it. Now does it have a lot of things that we have to improve on, that OpenAI does? Absolutely. And I think the case is the same with all the panelists up here, but it's here to stay, and it's going to accelerate our understanding, and it's going to have one of the biggest like profound impacts on this industry. And I cannot wait to kind of usher into the computerization of medicine because that's where we're headed with all of this.

So I'm going to present a cautionary note, like Stephen, I'm old and have lived through many aspects but on the other side of the world, which is on the machine learning world. And I've lived through a number of AI winters. And so while I completely agree with Sean that this technology is delivering an incredible amount of value. I also find a risk in over hyperbolic narrative because there has been multiple times where I've seen this going on where people make extravagant promises that are not borne out by the technology. And then there's an overreaction and the entire industry feels, I got my PhD at a time, when you couldn't say we're in a researcher, you have to say you were doing cognitive computing because it was just not respectable to be doing AI. So -- and that's because people made these extravagant promises that were not borne out by the technology. So I think the technology is transformative. I think it will bring a lot of value. It's not going to be this over bullet for drug discovery. It's not going to transform everything and you're going to get 1,000 drugs to the clinic in 3 years, which are some of occurring that I've heard. So I think you have to be nuanced and careful on how you do narrative so that you don't get this overreaction in the other direction.

From the R&D sharing economy corner. It does help with sales because people want to try this stuff out, and it's often easier to try it with an expert service provider than trying to spin up a whole. It's less the bite and spinning it up yourself. And so I think it gets people curious, gets big pharma and even small biotechs curious, and that's actually, I think, a help, so...

What do you think helped? It'd be more real to people? Is that like clinical trial...

Yes, data in their area, 100% biopharma trades on data. Like the biggest thing for us, we added 59 new contracts last year -- cell programs last year with customers and 1/3 of them -- more than 1/3 were in pharma. And until we had data in various modal -- first data in AAV, first data in cell therapy, we couldn't get deals, right? Like people just -- it is an industry that trades on show me the data. So yes, 100%. That's what we did.

I totally agree with that. And I think like -- that's -- I mean that's what we're seeing is like we're actually seeing the data, the proof is in the pudding, like generative AI is starting to create antibodies from scratch that are better. And I think like you use that and you kind of like go out 5 to 10 years, I think you can start to see the huge impact of what this is going to have for our industry.

Well, great. We are out of time. I want to thank you guys again for a really engaging panel.

Thank you.

Thank you.

Thank you.

Thank you.