Hey, good morning, everyone. My name is Siyuan Chen. I am the CTO on Twist Bio. I'm an oligochemist by training. Actually, I was the first employee on the R&D side at Twist. I've been working in this company for 13 years, a long time. I'm really excited to be here because I don't think I've ever talked to a large group of investors in this setting. I'm a little bit nervous, but I'm very excited to show you what we're able to do. I think you guys all just went through the tour, right? The production tour. I'm sure you have seen a lot of really interesting, a lot of amazing stuff there. I want to say a lot of magic we do, really started with the silicon chip. I think you have all seen a silicon chip, which is 96- well plate size.
Instead of having 96 wells, we actually have 10,000 wells on it. We call them clusters. Within each cluster, we can make 121 oligos. That's enough oligo to make one gene. On the chip, we're able to make 10,000 genes. On the photo chip, we're also able to make about 1 million oligos. We have multiple writers. The writers can run actually really, really fast. I'll talk a little bit more about it later. That give us a capacity of making 32 million oligos per day capacity, enables a lot of different applications. We also do the DNA synthesis base by base, like using phosphoramidite. We have done so much optimization to really improve the efficiency on every single layer.
Right now, we're actually able to do synthesis all the way to 500 base pair, so we call them 500- mer, which is actually quite amazing if you really think about it, because, as I said, I was an oligochemist. I remember when I was doing work in graduate school, trying to making 30- mer on a chip. That's enough length for hybridization, like for DNA microarray. That's quite amazing already. You're able to do 30- mer and the quality good enough to do hybridization. That's good. If you're able to do 80- mer, 100- mer, everybody look up to you saying that you can do 80- mer in a consistent basis. That's quite amazing, but right now, we actually do 500- mer in a consistent basis on a production setting. Also error rate, that's industry-leading.
I think we're able to get to one in 2,500 base pair, one in 3,000 base pair error rate. That's really unheard of compared to the traditional synthesizer, you get a five in 500 base pair. That's pretty much as good as you can go. That's our foundational platform, and on the platform, we really built up a lot of DNA product. Oligos, oligo pool, gene fragments, clonal genes. Also in the last couple of years, instead of printing A, T, C, G, four different bases, we're actually able to add in modified bases to make some siRNA antisense oligonucleotide. You're actually going to hear a talk from Dillon from GSK later today talking about how they use our platform to enable high-throughput siRNA screening.
On top of it, actually, on top of the DNA layer, we also built up a protein layer where we can do different type of antibody discovery work and antibody production and antibody characterization. We have this amazing platform where we can make DNA in a scale and a very fastest speed. We really built up a very impressive NPI machine on top of it. I want to show you where we were back in 2021. This is essentially the product we had back in five years ago. We have syn bio product, we have NGS product. On the syn bio side, we have oligo pool, we have variant library, we can make gene fragments, we can make clonal genes. On the NGS side, we have exome panels, we have some custom panels, we had a couple of library preparations. They're all good product.
We are still selling those product very strong these days. At the same time, I want to say the footprint, the product lines is somewhat limited, and I want to show you in the last five years what we have done in terms of expanding our product roadmap here, our product line. This is where we are right now in 2026. I'm not going to go over all the product here, but I want to highlight a few things we can do. On the gene side, we have clonal genes. Two and a half years ago, we launched the Express Genes, which offers industry-leading turnaround time for clonal gene product. A few weeks ago, we actually announced ultra complex gene and SynBioBeta.
That really leveraging the 500 base pair synthesis capability I mentioned a couple of slides ago, so that we don't have to really worry too much about DNA secondary structure. We're just really going to make the DNA chemically and stitch only a few pieces together to make ultra complex gene. That serves as a backbone for a very complex mRNA and help us get in the world of nucleic acid therapeutics. You're going to hear more from Paddy later today talking about that part. At the same time, on the synbio side, actually, we built up a protein layer on top of it, where we can do antibody discovery in vitro, in vivo, and we also use AI tools to do in silico discovery work.
At the same time, we built up a very impressive antibody expression system and antibody characterization, generate lots of data to help our customers to do AI model training, refinement, and help their AI-enabled drug discovery work. Colby is going to talk more about it, share more insights on the AI side. On the NGS side, we expanded our panel product, lots of panel product. We put a lot of emphasis on MRD, molecular residual disease. Which we really think that area going to ramp up very, very quickly. Jimmy going to cover that part later today. On the library preparation side, we had a couple of library preparation.
We added quite a few more new product, found some generic library preparation, some standard fragmentation, and also some more specific ones focusing on cfDNA, focus on FlexPrep, to really enable microarray conversion [from two to NGS]. That's actually really enabled by the enzyme engineering capability we developed in the last few years, leveraging our synthetic DNA synthesis capability. It's a really nice product. What makes it even more exciting is I think that's actually just really the beginning of what we can do. If you really look into all the application, can be enabled by oligos, by DNA, in terms of oligo number, mass, length, in different areas. We can see there's a lot of things we can do, which is we just need to choose which one we want to go as we continue to move forward.
Yeah, this is the table just to reference the way how we perceive what DNA can do for different applications. That's enough about NPI machine. We're going to talk a lot more about it throughout the day. What actually I want to spend the rest of my talk here to share with you a little bit more about how we do operational excellence, because that's actually a really important part. We spend a lot of time on it. We're really excited about it, and that's also a part might be a little bit underappreciated by people from the outside world. I want to share a little bit more what we do. We always come back to the silicon chip, right? That's really our foundational platform. We do pretty much everything on top of it.
I want to say the chip itself is actually not static. It's actually, we continue to iterate and improve this platform. A few things I want to share. Number one, we're able to maintain really, really good error rate in the last couple of years. Like one in 3,000 base pair. Sometimes the best run we ever seen was one in 4,000 base pair, which is super amazing. We also improved the consistency of the DNA synthesis quite a bit in the last couple of years. Believe it or not, actually, the DNA synthesis can be pretty variable, can fluctuate quite a lot. There's even a lot of seasonality in DNA synthesis. Really think about it, just even if in a rainy day, the moisture can be high, that's going to affect your DNA coupling.
Like even the rush hour, the ozone from the rush hour can actually affect quality of the oligos. Sometimes for the vendor, we buy all the bulk chemical from the vendor. They might store the chemical outside in the hot summer. The quality is still going to be good, but it's actually going to have impact on the oligo quality. We have done a lot of work, actually, to just to really tease out all the details, trying to improve the performance. I want to say right now, if we look at longitudinal data, for the oligo synthesis has been very flat, month by month, and across all the machines that we have here. Which is something probably cannot be said by other competitor when they have hundreds of machines running plates over plates.
One common comments we hear from a customer is they struggle with the current provider because they get lots of batch-to-batch reproducibility issues, things like that. That's not something we have to worry about because of all the work we have done on the synthesis improvements. In the same time, we continue to reduce the cost of the synthesis. As you see, in the last three years, from 2023 to 2026, we're able to reduce the synthesis cost by 60%. A lot of cost reduction come from the use of less solvents. Right? As you see in the middle, when back in 2023, took us about 51 L of chemical to make a million, 100-mer. That's quite amazing already because that's a million different sequences, 51 L of chemical. We did a calculation back in the days. It's actually 99.8% reduction compared to the column-based synthesizer.
Which is we're talking about three orders of magnitude lower than what people normally use. Yet, in 2026, we're able to reduce that number to 14 L for 1 million oligos. What does that mean, 14 L ? If you really think about it. Look at this bottle. This is a 500 mL bottle. A bottle of solvent like this can be used to make 35,000 oligos. Right? I think that's the scale we're talking about here. I think for the people who are around DNA synthesis, a bottle of chemical probably normally used to make a few oligos, but here, we're able to pack in 35,000 oligos into a small bottle of solvents here. More impressively, I think we have done a lot of work to reduce the turnaround time.
In the first half of 2023, took us 26 hours to make 1 million, 100-nucleotide oligos, which was really fast at that time. When we launched Express Genes, we tried to really look into everywhere, every way, every place we can shrink the time. On the writer side, we were able to reduce it from 26 hours to 13 hours, and at that time, I was joking, "Oh, that's super fast already. I don't think there's much we can do to make it even faster." Yet in 2026, we're able to shrink it down to seven hours to make 100-mer, and that's actually a 73% reduction in turnaround time. Not only we get 37% reduction in turnaround time, which actually means we have 4x increase in capacity.
We have the same number of writers compared to three years ago, but our oligo sensor capacity actually increased by 4x with the same number of writers. That actually enabled a lot of new applications, new product. For example, I think I talked about the ultra complex gene, which need a 500- nucleotide oligos. If we took the 2023 chemistry, it would take us 5.5 days to make a 500- mer on the writer, and that's going to take up a lot of capacity, and also, the product is not going to be competitive because from the get-go, you spend almost a week just making the oligos. Nowadays, we can actually do it in less than 30 hours for the 500- mer, which really enabled ultra complex gene product.
At the same time, we built up enough enormous amount capacity, 32 million oligos per day capacity, really support the needs for MRD customers because they all come in with the personalized panels. Another thing we do, it's automation. We actually take automation very seriously. We try to do automation from start to finish. When you go through the production tour, I'm sure you see a lot of automation. We actually use a lot of Hamilton, right? That's the liquid handler to move liquid, pipette liquid, move plates around. In a lot of applications, in scenes like Hamilton works well enough, right? Just load the deck, load the tips, load the sample, and in an hour or so, come back with what you need to do.
That works well for a lot of applications, but not good enough for the gene production because the gene production has many steps, has 20+ steps in the process there. If we were to use standalone workstations like Hamilton, the issue we're running into is it takes operator time to set up the deck, load the tips, load the plates, and run it, and once the run is done, have to move the plate from one machine to the other machine, do it again. It's actually quite labor intensive. That's created a lot of bottlenecks in our manufacturing process. What we did here is we identified all the key bottlenecks in our production process. We built up integrated automation, integrated system to enable all the bottleneck steps here.
Those are four major systems we built from oligo fragmentation system to take the rider oligos to gene fragments, and then the second one to clone the fragments, then we plate them, we pick all the colonies. We have to pick millions of colonies to support our demand, and then once we pick the colonies, all the colony goes through the Next-Generation Sequencing sample prep system. We actually process about 10 million samples per year on gene facility, which I believe we're probably more than the highest volume in terms of the sample volume for the NGS. I don't think anyone else in the world can match the scale, the number of samples we process on sequencing every year. Once you identify the perfect clone, we can go through a plasmid preparation to get purified plasmid and they're ready to send to customers.
That's essentially how we enable the scaling with more and more integrated system, right. As you see right now, we have a total of 20 integrated system on our production floor, and some of them you already see on the production to enable genes and proteins. We have quite a few systems in San Francisco as well to support our NGS product. That number was actually much, much lower just a couple of years ago. We only have seven systems. We continue to develop the tool, implement the tool, continue to improve our production by having more and more integrated solutions. One thing I really want to highlight, it's when we build tools like this, we always think through scalability and trying to make it future-proof as much as we can.
When we always look into when we brought a new product, new process, we really want to make sure multiple process, multiple product line can run the same machines. Back to the ultra complex gene, which the workflow is somewhat different from what we normally do for standard genes, but we're able to leverage all the systems you see on the left side, and with minimal addition of equipment. That's how we want make it future-proof and all future compatible. At the same time, with auto automation, we're able to improve our capacity, improve our throughput, and also with lower footprint, right? When you walk into the Gene Lab, the Gene Lab 1, that's actually the footprint of the Gene Lab 1. We used to have two assembly lines as you see on the right side, those two green blocks.
That's all the individual automation back in the days. That's operators basically do work on Automation 1, finish the work, move to the Station 2. As they work down the aisle, they go from oligos to gene fragments. That's what we used to do. It's automated, but we still need three to four operators running through the process. It takes about 9- 10 hours to go through the process because it's automated, but you do need pretty much all hands on time to make it happen. Over 10 hours, you can make 12 plates. In the last year, we actually put in two integrated system, the two solid green box on the right side.
The footprint is 1/5 of what we have for the gene lab for the original workstations, but we're able to, in these cases, operators only have to come in, a couple of people loading the plate, loading the tips. They come back in six to eight hours with 16 plates ready to go. They don't have to worry too much about it. They can actually do other work during the day. Now with the integrated system, we're able to manufacture 2x the fragments with 1/5 of the space. Now, actually, we open up the Gene Lab 1. We're doing more specialty work. We can use it for cell-free work. We can use it for mRNA work in the future. That's how we ensure scalability and sustainability as we continue to grow. In Twist, we always go crazy about speed.
I feel like we're always in pursuit of speed here. I think I already talked about the story about how we reduce the time for oligosynthesis from 26 hours to seven hours in the last three years. We did something pretty similar for gene product. I remember many, many years ago when we launched a gene product, took us 30+ days to make a gene. It's a very slow process, and we refined the process. We're able to make standard genes up to 5 KB in 10- 15 business days, and when we launch Express Genes, we're able to really shrink it from 10- 15 days to four to seven business days, which is faster than anyone else in the world. If you want things to be even faster, we can go with gene fragments, like as two to four business days, super high quality.
I think James just gave an example, the one-day fragments we made for Mike Wiley, actually someone we worked with a few years ago during the Ebola outbreak on the NGS. We're able to deliver genes in less than one day. Actually, that gene spent more time on the FedEx truck than in our facility. That's just how fast we are, how crazy we are when it comes to keep continuing to improve the turnaround time. We're doing something pretty similar on the IgG side as well, right? When we launched our product, 20-25 business days, good, but not fast enough. Able to reduce it to 10-15 business days as of today. If you want things to be even faster, we can do them in cell-free manner, and you can get your antibodies in five business days.
I think on the operational excellence side, we look at speed, because we always want to go faster. We also try to reduce the turnaround times that with the same number of people, we can do more work. We always look for opportunity where we can save money, reduce the cost, and also ways for us to improve the capacity, improve the throughput. This basically shows you what we have done in the last three years. Actually, even in the last a year and a half, we have 46 tracked projects to improve our tracked CPI, continuous process improvements, to improve our operational excellence, resulting in tens of million dollars of saving in the last two to three years. I am not going to be able to go over all the details about these 46 projects. That is just too many.
I do want to highlight one project we did on the panel side. We always make, on the NGS, I, we make panels. We're always the best when it comes to making panels. The highest quality, fastest turnaround time. We can make panels in two to three weeks, which is much faster than other people, which could be easily double our time. Starting a couple of years ago, actually, we saw MRD. We definitely see that personalized MRD is going to be a big trend. People are going to do more panels. It's going to be all personalized. There's going to be lots of them, right. Just to find out understanding, we're like, "Hey, customers really need the panels to be incredibly fast because MRD patients cannot wait." The window for the test is very short.
We want to make the panel as fast as possible, ideally less than a week. We also need to build a very highly efficient process because we need to be able to make hundreds, thousands of millions of panels, because that's how many cancer patients we're talking about. We cannot compromise the quality. We still have to maintain the highest possible quality because all the patients deserve the best panel we can make for their cancer detection. We know that's what we need to do. We planned a couple of years ahead of time saying, "Hey, we see the surge coming. It's not there yet, but we need to be ready." We did a lot of planning, a lot of preparation. We know we need to put in fully walk-away automation system to make the panels.
We need to build in the software to really make this happen, the integration of the soft MES and the hardware. It's all up to execution. I think one thing we do really, really well in this company is execution. At the end of the day, with the fixed headcount for the panel production, we're able to increase our capacity by 10x and actually reduce the turnaround time very significantly. From two to three weeks to make a panel, to less than five days to make a panel. At the same time, we're actually able to take the learning we have to make MRD panels, and expand it for all the panels, actually resulting in more than $2 million saving in the consumables when it comes to panels. That's actually the year-one saving.
The number is only getting bigger and bigger as we continue to scale our operation. I think this is my last slide here. I think just want to give you a taste of how we see the world, when it comes to NPI, it comes to operational excellence. I'm really incredibly excited about what we're able to do. I'm happy to talk to all of you offline. I'm going to pass the mic back to Emily here.
Thank you very much, Siyuan. Much appreciated. Next, we are going to have a fireside chat with Dr. Frances Arnold. It's my great pleasure to bring Frances here. I have my note cards. I feel like Alex Trebek on a game show. We have a runner. There's someone with a microphone-
Right.
Thank you, in the room. I have my own questions, but if you have a burning question, please do not hesitate to get in. To start, thank you, Frances. I remember when we were talking about Twist when we were a baby startup company. We were thinking about who can you bring in the SAB. We need the best, and I told the board, you don't know that, but I'll tell you now, we need Frances before she wins a Nobel Prize because we'll never get her after. You graciously accepted. Maybe my first question, we'll start there is, we had a number of SAB meetings when we were just at the beginning. A lot of startups fail. Did you think we were going to make it, or did you think we were crazy, it was never going to work?
I thought it was crazy, but that you were going to make it. The need was so great, and the ideas were clearly the right ones, right? To miniaturize and go with the silicon. I'm an engineer by training, so this is deep engineering and not appreciated by many biologists. I thought you were exactly the right person to make that happen, and you did it.
We did it. Well, if it's good, it's the team. If it's bad, it's me. At Twist, I'm only in charge of the bad stuff. We just did the tour. What do you think? What surprised you, in good or bad?
Well, I have to say I'm blown away because I moved away from the SAB in the early days. I had to deal with a whole bunch of other things at that time, so I didn't follow in detail. This is the first time I've actually visited the manufacturing. I love it. I have to say, I love it. The attention to execution. Although you were always attentive from the very beginning. You put in quality control as number one, careful engineering as number one. I'm not surprised, but it's fantastic because I use these products all the time in my research, and we love it. We absolutely love it. I'm not surprised to see how well you've been able to scale this.
That's great. You're an expert in enzyme. I think a lot of your career has been focused on enzyme. I know you have to catch a plane after this, so you'll miss our part on enzyme engineering. Just tell us why enzyme are so important from your perspective.
I've been working on enzymes forever, and these are those remarkable molecular machines that convert really cheap materials like carbon dioxide and sunlight into complex chemicals like trees or you and me. All the chemistry of the biological world is done using these machines. Now we're coming into a period of synthetic biology. It's not just all about healthcare, folks. It's about everything. Biology can make virtually anything. It's just bringing the cost down. It's bringing the cost down and opening up, expanding the possibilities for biology. Biology makes you and does a pretty limited set of chemistries. Evolution is an algorithm that can go out well beyond.
I see now, enabled by what you're doing and AI, a combination, as enabling biology to do any chemistry you want, to make any pharmaceutical, to make any fuel, to make textiles, to make all sorts of things we need in our daily lives that are done using dirty chemistry today, or not even done at all. I think we're at an inflection point, really, that's more exciting even than it was 12 years ago.
I totally agree. Yeah. Enzymes are proteins, right? There are non-therapeutic protein in the two big categories. There's also the therapeutic protein side of it. You're an academic. You're in the SAB. You're founder of a number of companies. You're on the board of Alphabet. You have, I think, a unique view into protein engineering and also as it relates to therapeutics. AI is changing that field. One of the questions, and I'm sure a lot of investors in the room have the question, is around AI-driven drug discovery. Is it a flash in the pan, or is it here to stay?
Well, what do you see? What are you seeing? Because you're seeing the same thing I'm seeing, and I want to hear it from your mouth.
Well, I'm going to say what I see. What I see is we're going to need a bigger boat in terms of capacity. I think that what Twist is, there's a first pass of model building, where people need a lot of sequences expressed in protein, tested against a number of targets to build the model. We find a model, then you have to turn the crank. Turning the crank needs a lot more DNA.
I think what we're seeing is AI is going to be the first pass for drug discovery. In vivo and in vitro are still going to be needed. AI has a huge advantage is that it's fast. It's much faster than having to either inoculate a mouse. You have to wait for the mouse to be your drug company or you have to do phage display or yeast display, and that takes days. That's what we're seeing is AI as the first pass. Does that jive with what you're saying?
Of course, because enzymes are more complicated than the binding proteins used in therapeutics, it's the next generation, right? AI is still not good enough unless you combine it with these optimization methods for which I won the Nobel Prize. Imagine that you can combine AI to get you a good starting point for chemistry. You combine that with iterative optimization, all of which uses synthetic DNA in an active learning cycle. You just press the button, and I think in the next five years, we'll be able to genetically encode any chemistry. I'm super excited about this, and I know it's not a flash in the pan. I write the checks for that at Alphabet. They're really big checks.
It's not a flash in the pan because whatever you can't do today, that won't be true tomorrow, and we're just going to get better at it.
Do you think it's going to make DNA obsolete, or do you think it's going to create more demand for DNA because they're going to have to try more things?
That's a really good question. If we get perfect at AI, you just make one gene and you're done. That is completely unrealistic. There's so many specifications that we don't even know to make on the DNA, how it cures a disease, how it catalyzes a reaction in a particular laundry detergent formulation. There's all sorts of things that we can't write down as specifications that only become real when you translate the DNA into the sequence that you compose using AI into the real world. I think you said it this morning, DNA is the point at which you translate your computation into the real world. That is the physical manifestation of your computation. I think it always will be, because it's not efficient just to go and make proteins synthetically. Nature does it by DNA, and she does it for sugar.
I do it for sugar, too.
A lot of sugar.
Lots of sugar. If there are any questions, raise your hand and we have a runner that will bring you questions. The last nine years, actually, we had to write down, we had a number of projects together. Early on, we did the membrane protein chimeras, and then we did machine learning design channelrhodopsin optogenetics tools. Very recently, we did a carbene transferases project. As a user of DNA, how do you choose a provider? You're using Twist a lot, but there are other providers. What's the criteria for you to make a decision of where to go, or do you leave that to your team?
Well, I come from academics. We never have enough money. Price is very important. That's why I yank on your chain and say, "Hey, I want a better price," but I'll give something in return. Yes, price matters a lot for the academics, but I also sit on the industrial side, and their time and quality are the key things. They're different metrics and I have to say, I'm pretty impressed that you can meet all of those metrics.
Yeah, a lot of companies, between the speed, quality, or price, usually you have to choose two.
Any two. With Twist, we pride ourselves in you get all three.
That will be important because graduate students are relatively cheap. They used to be almost free, and so manipulations, that wasn't important. Now they've become a little bit more expensive, and so we actually do a little bit of calculation over the trade-off. Do we want the whole gene or do we do all that ourselves? In industry, especially in this new AI era, turnaround time is really critical. As we move into active learning test build design cycles, the ability to turn that around and generate the key data will mark who wins in this race. Believe me, it is a big race. There's a huge amount of interest in biology as a manifestation of AI. Not just therapeutics, as I say, it's all of chemistry. It's going to be a race who can do this.
That's why we love speed. Yeah. There's a question. Great. Puneet.
Great. Dr. Arnold. Thanks for the question, and great to see you here. The question is really about there are a lot of assumptions being put forth in AI in terms of speed to therapeutics, speed to drug discovery. Can you talk a little bit about in terms of scaling, as you talked about briefly about chemists' role is going to change a bit, too. Can you talk about the scaling we're going to see? There's an assumption that we're hearing in the market that in 10 years, we can see a number of therapeutics coming to market that can actually resolve a number of diseases. That's a big assumption being put out there.
Disease and biology is complex. Can you talk about the scale that needs to happen, the hiccups that can potentially happen, and then how long is this versus 10 years?
I sit on both sides. Demis Hassabis says we're going to cure all diseases, but he doesn't talk to the FDA. You can't just go and test these things willy-nilly. Yes, AI might cure some number of diseases, but it's all going to be at the rate at which drugs can be developed. The pharma people are much more realistic, but maybe not quite as visionary. What's going to happen? A lot of it depends on what happens on the regulatory side and on clinical trials. How do we understand the efficacy of these AI-designed drugs?
Thank you. Any other question in the room? I'll keep going, and if you have a question, please do raise your hand. As a scientist, I talk to a lot of scientists, and we ask people about protein engineering or
Any topic almost, you ask three people how to do it, you get five answers. Right? I'm always shocked that there's not more standardization, we're trying to turn into a competitive advantage by saying, "Oh, we'll do whatever you want," why do you think that is that there's not more standardization? Everybody wants to do something different.
Hell if I know. Well, scientists are funny ducks. They like to put their own stamp on it. That's one thing. There might be one great method out there. That doesn't mean everybody piles in to do it, especially with a complex problem like protein engineering. Every protein is different, and every landscape, we call them landscapes, how you go and explore different sequences is different. How you measure everything is different. There's a great deal of bespoke engineering that goes into it, which really tends to push people into their own methodologies. That said, I do think, Emily, that there will be a push the button that maybe not be optimal for each problem, but will be optimal across most protein engineering tasks, and that will involve machine learning, active learning cycles with machine learning.
That project we did with you way back then with the optogenetics, that was the first time machine learning was applied. In fact, we developed those methods in 2012, 2013, and did that collaboration with you to demonstrate how you could engineer new properties, and that was well before the ChatGPT era. Yeah.
Well, speaking of ChatGPT, I have a personal question. When you use an LLM, ChatGPT, it's all about the prompt, asking the right question.
Obviously, you're massively accomplished. How were you able to ask the right question? How do you think you got here? What made you get here?
I think it was desperation. If you're pushed hard enough. I came out of an engineering background, and I jumped into protein engineering when it was a brand-new field. The people who were doing it, and I'm sure you have had exactly the same experience, the people who were doing a little bit of protein engineering were from structural biology, and their whole mindset was you had to get a crystal structure of this very complex molecule first, which often people couldn't even get that, in order to go in and then rationally design everything. I came in and I said, "Well, I won't get tenure or I'll die first before that happens." I had to come up with some very different engineering mindset, which was look to the best engineer on the planet, and that's called evolution.
To me, it was totally obvious, but to the field, it was completely non-obvious. I think in DNA synthesis, you experienced the same thing. Everybody was doing it the same terrible way. Terrible way, and there's no way you could scale that. You came in as a chemist engineer and said, "No, we have to completely rethink that." Isn't that the case?
Yeah. Being a contrarian goes a long way, right? I always say, one, to the team is because what we do is incredibly hard, and sometimes it's like, "Oh my God, it's so hard." I'm like, "Good. Good. If it was easy, every idiot would be doing it, right?
Exactly.
Number two, when there's a plan, "Okay, let's do this," I always ask why. Why this way?
The worst answer for me is like, "Oh, because everybody else is doing it.
Right.
I try, no. We're not doing that because everybody else, they have more resources than us. They have more experience. They have better channels, better capital access. How can we beat them if we do the same? Totally agree.
You also have to have a vision of when you enable this capability, what becomes available and how do you capture that and capture at least some of the value of that?
No. Yeah, totally agree. That's why I think it's very helpful to be technical. When I started Twist, I kind of was downplaying my Ph.D. because they're like, "Oh, you don't know business." I think you can learn business easier than learning the technology. When you talk to customers, if you don't understand what they're doing, when you don't even know your product, how can you be effective? I think that has been very useful, being able to, you drop me in any accounts, and I can sell any products that we have.
What I love is what the technology has enabled that we could not do before. In my work, some of the same things happened, and that's how I chose problems was, what can I demonstrate that you just couldn't do with rational design? A good example is how do you make an enzyme work in a laundry machine? That was a big market for enzymes. How do you make it be happy? What self-respecting natural enzyme wants to work in your laundry machine with bleach and surfactants? There's no way you could design that. It had to come from directed evolution, and you can't go out to nature and find that. You had to have some methodology, and you've done much of that, right? My methods enabled people to do a whole bunch of much more important things than make laundry enzymes.
Your technology has also enabled the whole synthetic biology industry, I would say, to do things well beyond what even you could imagine.
Yeah. No, that's true. Any questions in the room? Okay. All right.
Thank you for your time, Dr. Arnold. Curious, there's a variety of models that exist today, AlphaFold, Boltz, and others that I think are adopted by pharma. Would be curious for your perspective on the utility of those models in drug design today versus what might be required from a new data generation standpoint to ultimately get to the point of, maybe not quite to the point, but a sort of push button get drug, or get closer to that point, rather than the models that have been largely built on third-party existing data.
I'm very excited about the models. I love that Demis won and David Baker won the Nobel Prize for understanding protein folding and then design of proteins. The bottom line is that structure is not function. Right? That being able to predict the structure is a game. It's like winning at chess, and you have a good metric for that. What we want in the synthetic biology community is something that does something. The models just can't capture that right now, but I think they will. They'll start beginning to capture that as we get the right kind of data, which we don't have. Even though that problem has not been solved, we are getting close to getting things that can solve it through further experimentation.
This is why it's a beautiful time for making DNA, because we're close enough that we just need a whole lot more experiments in order to learn what it really takes to make something that's useful.
Question over there. Yeah.
Hey. To that point, I guess, if a lot of the open source tools are protein structure today, and a lot of the drug developers are incentivized to have more siloed data sets in regards to function, how do you see this playing out over the next decade? Is that going to be the path forward, that each developer will keep some of that data in-house? Is that going to be a rate limiting factor on the field in general, or do you think these open source tools will move towards function and more downstream practicality?
I think that's a really good question. I don't know how to answer that because I don't know how much data it will take. There are those who argue that enough data on your particular system is all you need. You don't need a world model that works across all modalities. To me, as an experimentalist, that makes more sense because I know how bespoke every protein is. On the other hand, if Demis is right, you learn across all proteins, and you just have a model that does anything. He doesn't know anything about proteins, so to a first approximation. There's no quotes given here, right?
No.
This is Chatham House Rule. I love Demis. I don't know who's going to be right, but I do know. I'm on the board of Generate: Biomedicines, so it went public about six weeks ago, AI-produced antibodies. We integrate it with a lot of experiments to get to the right developable drugs.
I think there was a question in the front. No? Questions? All right. Maybe the last question for me. What if you have to start again today? Suspend disbelief, you're back, going to universities. If you had to do it all again now with the AI tools, with the ability to have the wet lab outsourced, even maybe at Twist, either at Caltech or not, if you had to start again today, where? What would you work and work on, and how would you do it?
Two answers. I want to ask you that question. I love proteins. Protein engineering is not solved. You could jump in at the place it is and still do really important work. Enzymes are not solved. I am a student of evolution. Evolution works at all scales. Why not apply some of the same ideas to tens of molecules, to whole cells, to organisms, to ecosystems? It's the same design process to design anything in biology. That means whole new scales of DNA, right? Not just one gene, but a whole ecosystem of organisms. I do know that some of the most visionary people in the community are really thinking about that. How do we use AI, for example, to design whole genomes? That's happening. How do we use AI to design whole ecosystems?
That's going to require a hell of a lot of DNA.
I love it. I'll be.
What would you do?
I love DNA chemistry. I know to write DNA, read DNA, sell DNA. I think I was built to build Twist. If we had to start today, frankly, I think it would be very hard because when we started Twist 13 years ago, we knew that we had to raise $1 billion to get to exit velocity. Back then, 2013, you could do it. You could raise the first $600,000 seed round and then a $9 million A round, always telling everybody, "Look, at the end, it's going to take $1 billion." We could do it because capital was available. I don't know if we could do it today.
It's all being sucked up by SpaceX.
AI. Yeah. I'm glad we did it then because I think now it'll be tough. Yeah.
That's sad to hear, because the Anthropic are raising lots of money, and they're going to be buying all your DNA.
Yeah. We were able to do it, so it's good. Necessity is the mother of invention, I think.
Do you think one of those companies will buy Twist?
Now I'm being triggered. Hopefully they'll buy a lot of DNA. I'll always say that that's not our goal. Our goal is to ramp our revenue, and eventually, we'll get to buy Illumina and Thermo Fisher.
Okay.
That's not long-term guidance, by the way. This is America, where for sale every day, if there's enough zeros on the check, I will fly and wash a car.
Okay.
On that, thank you again so much. I know you have to catch a plane. I very much appreciate the effort. I very much appreciate the partnership along the years, and we can't wait to continue being your DNA provider or protein or RNA to enable the great work you do.
Thank you, and thanks to the whole team here. It's just marvelous what you've created.
Thank you. This one.
Thanks.
Thanks. All right. Oh, sorry. Back. Next, we're going to hear from Colby Souders, our Chief Scientific Officer. Colby is our own drug developer. I believe that at least four drugs that he had his hands in the discovery are in the clinical trial stage, and a few of them at phase III. Colby, take it away.
Thank you. Thank you, Emily. Pleasure to be here. Thank you, everybody, for coming and those of you making the trip. Hopefully, it's been a great day so far. Fitting to follow up that fireside chat with the AI topic, which I know is of particular interest for many. We heard a lot of great comments in that. Now what we're going to do here is dovetail that conversation into how we approach AI at Twist and how we solve this for other companies, and how we enable AI companies to scale and to fulfill their promise to the market that they are making. I'll dive into this in more detail and hopefully by the end of this you'll understand our position on the market, as well as what our solutions provide to solve that.
This is a slide you saw from Siyuan in the earlier slides. You can see, first of all, a lot of products that we make touch that AI-enabled drug discovery. I'll focus mostly on our IgG and antibody characterization, but also mention in the next few slides how all those other aspects that are touching that area provide those tools and what differentiates our platform in order to enable the AI drug discovery in IgG and antibody characterization. First point to that being, of course, we've heard a lot about DNA. For AI-enabled companies, it's not just a DNA product that some of them need. Many of them need protein. Most of them need data. They don't necessarily need a physical product from us. They need that characterization data. To enable that, we need to start from DNA.
All biological material starts at that point. We've built out that speed, scale, and quality to enable that. Of course, we've talked about speed. We know with higher speed, we can evaluate more candidates in less time. That really accelerates that design, build, test cycle for our AI customers and enables them to develop more therapeutics within the same amount of time. Those are all huge advantages. Of course, with scale, you need these very large data sets. I think what we've seen over the last two or three years is that folks would start with smaller data sets or they'd start with unstructured public data, and that wasn't good enough. They realized we need to do this at 10x, 100x, 1,000x what we were thinking two or three years ago. We've made that scale.
We've enabled not just our DNA, but also our protein solution side, our protein production, our data characterization delivery to match that scale because that's what those AI companies really care about. Of course, not only does this support more candidates, but it supports other modalities and different target classes as well. I'll mention this more toward the end, but we think about things in terms of binders and the field that's gone from mini- binders to VHHs to IgGs, but there's a lot of other modalities that these AI companies are thinking about. We're just at the very early stages of scratching the surface of all those other capabilities. Of course, you can have speed and scale, but honestly, it means nothing if you don't have quality.
Poor data, no matter how much you can make and how fast you can make it, will not be informative for a model, and it certainly will not develop a therapeutic. You can think of quality in a number of different ways and a lot of the traditional methods that you would measure quality, but honestly, we think about it in a couple of other ways, too. By providing flexibility in different formats or multiple production systems, now we're enabling our customers to develop their therapeutics in the context that they want. We heard a little bit on the fireside chat, one of the important things is there's so many different types of biologics. Each of them needs a different system, a different format.
If you don't provide customers with that flexibility to order it, the DNA, the protein, the data in the format that you need to fill your model or fill your therapeutic pipeline, then it means nothing. Again, we've got several different product lines. I'll talk a little bit about how our unique multiplex gene fragment and gene pool systems really enable new library development. That isn't just for scale, but again, for quality, because now you can start making combinations of different products in really unique ways that no other company can enable. That's a huge piece, not just of a product offering, but a quality because the design and the flexibility that a company has to make that design and then actually fulfill that is unparalleled with those products. Finally, that fit-for-purpose downstream use.
Again, I mentioned this, but the endpoint and the starting point is very important for these AI companies. They don't all want to follow this linear gene to protein to data. Maybe they want to come in at the protein. Maybe they want to just get data from us, we provide that flexibility so somebody can start and end on that train, if you want to say, at any point. That's very important. We've built this foundational system for DNA synthesis and protein production and data collection. The important thing, though, is you can't attack it from one side. Saying, "Okay, I can build scale, I can build speed, I can make product really fast." You also need that expertise. You need that expertise in the characterization to fulfill the AI/ML dream of data sets.
We've attacked it from both angles. By this, I mean we've been developing for over a decade now different in vivo and vitro antibody drug discovery platforms to fulfill these and to deliver the hundreds, maybe 1,000 therapeutic programs that we've done for partners. We need to establish all of those advanced characterization methods that all these AI/ML companies want on the back end of their data production systems. We've already built these out by having all the deep expertise in that full end-to-end antibody discovery platform. This has required a number of tools where we have things that support all of these in vivo systems, all the in vitro library discovery that we've done.
We've developed these tools, some being shown here, and all we have to do is now apply that to the scale and the speed that we've developed on the front end. That's been a very seamless process. Now we can deliver the binding evaluation, the affinity measurements, the developability characterization, those functional assays, the things that are fundamental and that these AI/ML companies want, but they want it in high quality but scalable data. That's really what differentiates us is that many companies try to approach this from one side or the other. Maybe a company only has expertise in scaling. Maybe a company only has expertise in deep characterization and antibody drug discovery, but we do both, and we merge those two things.
On the one side, we've got the deep expertise in antibody characterization, protein binding, things of that nature, but we're able to apply our scalable automation and operational excellence that Siyuan was talking about earlier. That's how we apply it. I'm going to get into a bit more detail on exactly the different types of workflows that these AI/ML companies are essentially ordering from us and crave in order to fill the pipeline of data that they need. Two main modalities here. One being the library-based workflow that is very wide, the other being the clonal sequence workflow that goes very deep. What do I mean by that? The library-wide workflow means a customer might come to us with tens of thousands or hundreds of thousands or maybe even more individual designs, and they say, "I want to test all of these designs.
Not sure how well my algorithm does, I want to test all of those from at least a very basic standpoint. Now I can narrow it down. At the end of that, you can see a lot of those actually feed into that clonal sequence workflow. What the clonal sequence workflow does is goes very deep. We apply all of those characterization assays that I was talking about to the proteins that we produce on the scale of not just hundreds, not just thousands, but maybe tens of thousands. Now between one side or the other, you can tackle a problem from one antibody if somebody orders it, all the way up to hundreds of thousands or millions.
Being able to utilize each of these workflows, apply them at the right time and in the right sequence, is very critical to the success of these AI/ML companies. I'll provide a couple deeper scientific examples of where we've applied this. First being on the library side. I'll have two examples here. One is where we had a customer who actually just wanted the DNA delivered. They wanted to do the screening in their lab, but they required our library technologies to enable them to even screen that in the first place. In the second example, it'll be a full end-to-end solution where we not only did the library production, but as I'm showing here, all of that panning and NGS output, the screening, the lead candidate selection.
Here in the first example, going from one Nobel laureate to the other, this one's with David Baker, and where we collaborated and did some work for his lab. In this particular example, the challenge was using the algorithms that his lab has developed, the ProteinMPNN tools and the RFdiffusion algorithms, to design binders, VHH's, nanobodies, however you might know them, to four different unique binding sites on proteins, and wanted to generate 9,000 unique sequences that were targeting those different sites. We used our Multiplex Gene Fragment technology to make all of those libraries, put them together, and then send them to the Baker Lab where he was able to then validate and pick out the leads using cryo-EM and different SPR techniques to measure and validate that those models were working.
It was very critical for us to be able to use our precise printing library technology and print those libraries the exact sequence that he wants to be able to deliver that to multiple targets with over 9,000 designs per target. Now, in the second example, this was a collaboration with Amazon Bio Discovery, and Memorial Sloan Kettering Cancer Center. In this particular challenge, Memorial Sloan Kettering designed over 300,000 unique sequences. This was very large because this was actually to an undrugged target. Very complex protein, something very complex that had never been targeted before in biology. We needed to use a very large library to interrogate this and figure out if we had any valuable binders. We did all the screening. We looked at 12 different populations, millions, tens of millions of NGS reads.
We found hundreds of individual clones that we then selected. We did all of this work with the library screening, not just the library production, but put them in yeast, did the protein selection, the screening, and we found a number of different, very interesting and very valuable targets. The great thing here is that they're coming back. Just like many AI/ ML companies, they realize we're probably not going to solve it on the first round. They're using the data from this first round of output to come back, optimize it, and do a second round. When we look on the clonal sequence side, there's a number of different applications here. Again, think people want a lot of data now. We need to go very deep to characterize this to train the models very precisely.
The first example here being a study where we partnered with Charlotte Deane's lab. She's a world-renowned expert as well, who designs open- source bioinformatic tools that are basically used universally across the antibody industry. Very impressive. The aim here was develop a tool to predict nanobody structure and properties. Her lab had done this for antibodies before and had published those tools and methods, nobody had ever done this for nanobodies or VHHs before. We worked with her lab in order to develop the wet lab data that validated and fed back into these algorithms to generate, we found a number of different, really unique properties that were very interesting that hadn't really been realized before for nanobodies. Sorry. Those have been now incorporated into that model.
In the second example here for the clonal workflow, this is again with Amazon Bio Discovery, this time with the Gray Lab. Here, what the idea was is we wanted to characterize all of the different models that were being published in the Amazon Bio Discovery website. What we were able to do was take 5,000 different designs across 50 different targets. It's a lot. It's a big study. We generated 70,000 data points across seven assays for this particular one. Again, went very, very deep. This enabled them to learn which of these algorithms were valuable to predict different properties. Not every algorithm was perfect for every property, but now we know which tools to apply to which problem within that platform, and we can learn and fine-tune those models now in multiple iterations of that cycle.
A very valuable data set, one that's being used for benchmarking most of the Amazon Bio Discovery tools, and that other folks that it's open access and people able to use. Suffice to say, we're fully aligned with Amazon's mission there to build this ecosystem of AI agents and be able to empower these scientific capabilities and make them accessible to many researchers, not just the largest companies or the well-funded AI/ ML companies that can design their own algorithms, but making them available to all. Those were some very detailed, specific scientific examples, but I think what might be most valuable here in this setting is talking about the customer journey. In particular, what do most of these companies do when they come to us with a problem, when they come to us with an AI/ ML workflow that they need to execute on?
Most of these companies start at what we would consider the model building stage. In this particular case, they might have a model that they've already developed similar to some of the other examples that I just gave, and they'll say, "Well, first, we want to do a pilot study with you. We want to know that the data we're getting is going to be valuable, that it's going to fit into our models, and it's structured the way that we need." That'll usually be on the order of tens to maybe hundreds of sequences, a fairly small study.
We'll do all the production and all the characterization and data delivery for those known sequences from the partner where they already maybe have data on that, and they're benchmarking us to say, "Okay, how accurate is your data compared to what we expect?" These are usually completed in just a few weeks, for a matter of $10,000-$100,000. Once we pass that pilot stage, we go into the first round of training. This is where we've got thousands of sequences, maybe tens of thousands of sequences for a single round, for a single target. Again, we will go through the full make and test cycle to deliver data. This is completed again in about the same timeframe. Slightly higher cost, usually less than $1 million to be able to feed into the model.
Now this is real-world data that they're using in their design algorithm. Again, now they need to learn. This will feed back in. Almost never are these models perfect the first time. Usually you'll see two to maybe five additional rounds that will provide fine-tuned training for these models. By that, we mean that it's predicting particular properties of the proteins that the company is interested in. Maybe they're assessing how well it fits also into other parameters. Once we go through this, typically we've established a really good, close long-term relationship with the partner. More importantly, we become embedded essentially into their make-test cycle. The interesting thing is that then most of these companies realize, "Okay, I need to build more foundational models." They'll say, "Well, I had this original model I came to you with.
I learned a lot from it, but now I want to predict a different property." Maybe I say, "Well, that went really well. I need more data. I learned from that process that I need more data to solve additional problems." Here is when we get into that library build process. We have tens of thousands or hundreds of thousands of sequences. Again, those are designed across a number of different proteins now. We're looking to build generalizable models. We want to say, "Okay, I'm not solving a problem for just this one property or this one protein, now I'm solving a problem that I can apply to a wide variety." These are much larger studies, but usually completed still in weeks, and for $100,000-$1 million sometimes. Sorry.
On the validation of this, that's very similar to what we saw. Again, we're going into that clonal sequence process where we're making, testing, all of those, and that's so very similar to what we just saw. Again, we go through multiple cycles of that to learn and feed that back in to build these new foundational models. The company has a great platform. They have foundational models. They've fine-tuned them, they've tested them, they've learned from them. Now they say, "Okay, we're ready for therapeutic discovery." Now they'll apply this to a set of targets. Usually not just 1 at a time, usually multiple. It'll be hundreds to thousands of candidates that they want to make and test.
It's not just making tens or dozens. When you're making a therapeutic, you don't want to take that much risk. It's better to make hundreds of thousands and overshoot it rather than undershoot it. We'll do the production. We have a second layer of characterization we'll do here. We'll get into functional characterization, so we can really tell if this was an effective therapeutic for that particular application. Again, completed within a matter of less than a month for under $1 million. Very effective drug discovery campaign. We definitely will take top hits into optimization models. Again, zero-shot discovery of these therapeutic candidates is not where it is today. Maybe in the future, but we're still a ways away from that.
Personally, I think we'll always want to do some optimization tinkering models whenever you're developing a final therapeutic. Again, we'll do hundreds to thousands of these in a similar cycle. Finally, once we're done with this optimization process, then the partner will nominate that lead candidate. They'll move those into efficacy testing, in animal models and tox models, at different CDMOs. Once those test, then they will enter the partner's therapeutic pipeline. The great thing is that these aren't just theoretical. We've been doing this and we've been doing it for a little while. Just recently, though, we've completed over 200,000 proteins expressed. Over 130,000 of those were assayed. This has generated 7 million data points for dozens of customers. Very impressive scale. What you see here is one example of that in the data output.
This is millions of dollars on a slide, basically, in data. A very impressive throughput and is critical for that training and process. Thank you. The amount of biophysical data and information that is available for this particular data set has enabled not only model building but also therapeutic development. That's the key here, is this is what we're talking about when we say we generate these large data sets not just to help enable them to build more models, but to enable therapeutic development in the future. Here's another great example of a multimillion-dollar project where, again, we're looking at 50 different targets here. Again, binding candidates, 50 different proteins. We're also benchmarking against control antibodies, so those are in red that you see here.
We're measuring the candidates that we are testing, the AI design candidates, against those benchmark candidates to find out which ones would be the best properties. Now, we compare all of this data together. Again, it's not just binders, it's not just biophysical properties, but it's all of this combined so that now you can select lead candidates. Here we see in this green box, you put all of this data together for somebody to, again, train a model, because a model will depend on good data and bad data, the data in the yellow and the red, but then when you're selecting the therapeutic candidate, now you can select from that green box. That really allows people to use the most out of all of this data that we're delivering.
The last example that I'll end on here is a really interesting one because this kind of illustrates the way the market goes. In the biology, every 20 years or so, a new model, a new method will emerge. Here, what we're looking at is actually a campaign where we ran in vivo, in vitro, and AI/ML all together. In this particular example, we have hits from each method. We took those all the way through functional characterization to find the best hits, and we found that hits from every method were valuable. Really the message here is that it's not that one method supplants and replaces the last. The traditional drug discovery methods are still very viable and very useful. AI is now just a third additional tool that we can add to the mix.
It's a very unique one and the newest in this series of in vivo hybridoma discovery originally, in vitro phage display discovery, and now AI/ML. The interesting thing from this one is actually the partner has told us that they've nominated the lead. They're coming back for more optimization on that, but the lead is actually from the AI/ML library, interestingly. That was their lead candidate, came from that design. It passed very well through the animal efficacy studies, very useful. We know this is working. We know this is a tool that people are going to continue to use, and we're very pleased to see that all three of these methods can work together in concert to find lead candidates. Again, kind of teased it a little bit at the beginning, but it doesn't end here.
Far we've talked a lot about monoclonal discovery, single candidates, but we see a wide future in different modalities. One of those being bispecifics. We put a lot of effort into this recently as well. You can see here, we don't just end at the multiplex gene fragments, like I showed for the Baker Lab study, but we also have the gene pools. Now being able to make a single gene of that length means now you can really uniquely pair different bispecifics together. It's a very unique thing to be able to do so that you can start, again, AI designing how these should come together. Now we can do that in that library method, but then the question is, okay, now when you want to go deep, when you want to have that characterization workflow, how do you do that in high throughput?
Because bispecifics are traditionally very hard to work with. That's where the B-Body platform comes in. We had the licensing of the B-Body platform from Invenra, and this works extremely well in that high throughput method to make hundreds or thousands of candidates very quickly and then characterize them very deeply and keep them in that format for your downstream manufacturing and CMC. This is where we think the AI/ML market is going next, or at least one of the ways that it's going, one of many that we will continue to support. We've been very proud to support it, and provide the data, the genes, the protein, wherever somebody needs to start and stop along the way, and we're very excited about the future of the industry as well and the different modalities that it's going to enable.
With that, I'll turn it back over to Emily.
Thank you, Colby. What does that mean in terms of business? I'll do a click here. Thank you. In terms of business, we mentioned that we were able to deliver a large number of antibodies that were assayed, large number of data points. To help illustrate, we wanted to share the revenues that we got from five different customers. It's not an exhaustive list. There is a mix of a frontier AI lab, a large pharma, AI- native biotech, a dry lab biotech. What you can see is that we have to meet customers where they are. They all have different colors because we have a full menu of product, and at different times, even the same company needs different product. That has been our strategy to become a leader in drug discovery is we had to, one, have the full menu.
If you want a hybridoma, we'll do a hybridoma. If you want a single-cell workflow, we have that, and so on and so on. Not only we have the full menu, but we have the best implementation of that full menu, and that is being seen here, and AI is turbocharging this. You can see in general it's up into the right, and customers are doing things differently. That is back to the key of what we provide is industrialization of what you want, very high throughput, but we're going to be flexible, and we're going to customize our solution to exactly what you want. In terms of the market, what does that mean? We are updating our belief in terms of what the market size are going to be, and please note that this is for 2030.
We think that the DNA market is going to be flat at remain at $2 billion. We're seeing that the antibody drug discovery is going to grow, and on top of it, in antibody discovery service, we have an additional half a billion dollar that is driven by AI-driven drug discovery. We think that the protein expression market is also going to grow. We think there is a big opportunity. Of note, the assumption that we have here for AI is that there's not one dominant model. This is an assumption that many models are going to win. The second assumption is that AI drug discovery is going to work. We heard from Colby, we think it's going to work. We have example of it going to work.
If it didn't work, then the market will not be as big. Last, we are very happy to share that the growth between FY 2025 and FY 2026 so far is in the triple-digit order growth. We've talked about AI in the context of drug discovery, but as a reminder, AI is broader than just drug discovery. Here are just three examples of work that has been published by customers. On the left, we have using AI and synthetic DNA to discover, engineer, develop CRISPR tools. In the middle, the same, you're leveraging AI for mRNA expression in cell engineering therapy, in therapy with mRNAs modality. The promoter, the enhancers, the terminal regions are very important. You have to engineer those regions. Nothing better than the combination of AI plus Twist DNA to be able to engineer that.
Last, on the right, we can leverage AI for protein engineering. We heard from Frances that she's doing it, others are doing it, and you'll hear from Paddy this afternoon, even our own experience in engineering our own enzymes and protein leveraging AI. We absolutely love AI-driven drug discovery. We think it's going to grow our market. It's going to be a great catalyst for Twist. We should not forget that actually the AI as a tool is useful in a much broader fashion. With that, now we're going to hear from our customers, and I will let Angela introduce those customers. We have a number of customers. I think you're going to like it.
Fantastic. Thank you, Emily. We've had some great discussion about our internal platform, and now we are going to hear directly from customers. We have some pre-recorded videos, and we have some customers here with us to present live and in person. Our first is a video coming from Joshua Meier, who is the co-founder of Chai Discovery, where he leads the development of AI-driven technologies to accelerate drug discovery and molecular engineering. If we could cue that video, that would be great.
Hello, everyone. I'm Joshua Meier, co-founder and CEO of Chai Discovery. Today, I'm excited to tell you a little bit about what we're working on at Chai and how we're leveraging experimental capabilities at Twist in order to fuel our journey. A bit of background about myself. I started my career at OpenAI back on the early team in the nonprofit days. We worked on GPT-1 and GPT-2 back then, showed some of the early scaling laws, and I realized that if the models were able to learn how to speak English and speak French, why wouldn't they someday be able to learn how to speak DNA and proteins?
I went to Facebook, of all places, where I trained the first language models, transformer language models for protein sequences, and then, before starting Chai, was Chief AI Officer at a company called Absci, another antibody pipeline company, where we also did a lot of great work with Twist. Another fun fact is I was actually one of the Twist beta users back in my academic days. I was working at the Broad Institute working on gene editing, and we actually tested some of the first oligo pools coming out of Twist. I've been working with the Twist family for a very long time and very fortunate that we're a continued partner of them now at Chai. At Chai Discovery, we're building a computer-aided design suite for molecules.
The big vision of what we're trying to do in this space is to generate antibodies that are ready to go as therapeutics. If you think about the process of making a drug, there's many steps that go into the process, even all the way from the pre-clinical stage. You need to get a molecule that binds to a target efficiently. For instance, what you can see here is we might take a drug target in purple. We'll have a specific portion on that drug target where we need the antibody to bind. Then actually what our models are doing here is these are diffusion models that are placing the atoms in 3D space, so that they actually bind the targets effectively. We need to do this in a specific way. We need to do this in a high-affinity way.
Even if we can get a binder to these targets, there's a big difference between making a binder and making a drug. These molecules also need to have a whole host of drug developability properties. When we produce the antibody, we need it to be specific to the target. We need it to be expressed with high yield. We need it to have low viscosity, so we can't have the antibodies self-interacting and binding to one another. They just need to bind the target. Then we also need them to be stable. Even if we can make antibodies that bind these targets and have these drug developability properties, we also need them to actually have therapeutic function. One of the really nice things about these models is that they reason at the atomic layer.
These aren't like the NLP models that I was working on at OpenAI that reason in English or reason in French. These models actually reason in atoms. That's allowed us to come up with very specific designs. For instance, you can see here a peptide MHC complex where we would like to get a design that is specific to a single point mutant on a cancer. It turns out that now the models are able to do things like this. If we can bring all these challenges together, the outcome will be generating drug-like antibodies, including to hard targets which are difficult to go after with traditional methods. Really opening up the surface area for the kinds of molecules that we can make and the kinds of targets where we can apply them. How does Twist fit into this picture?
If we look at how the field has evolved over literally the past year, the success rates of AI methods in molecular design have really gone through the roof. Literally a year ago, back in May of 2025, the state of the art for antibody design was a 0.1% success rate, meaning one in 1,000 of the antibodies that you would make would actually bind in the lab. One of the really exciting things that we've seen with Chai is that if you look at the Chai-2 model that we published in June of last year, we were able to go from that 0.1% success rate to about a 16% success rate. Meaning if I make 1,000 antibodies, now 160 of them are going to bind. Really, an over 100-fold improvement in the number of designs that are binding the target.
In order to make this happen, there's both a ton of compute power that we need in order to train our models, but then, of course, also the data sets in order to train the models as well. This is where a company like Twist can really come in. I think Twist really has this potential to provide a ton of the data needed in this area. The technology is, again, a great fit, where you can start from the beginning of actually synthesizing the DNA. If you think about what happens when we're building out a training data set at Chai, you might have a certain sequence that you want to design. We might use our model, for instance, to design an antibody molecule. We can take that protein and we can think about what the DNA is. Twist, of course, great technology in DNA synthesis.
We can either make gene fragments, so each one of those sequences we can order those independently, or we can actually even bring this together into larger oligo pools and then synthesize many molecules at the same time, so we can scale data sets that way. The other thing that we would like to do is if you look at the way that antibodies are conventionally discovered, you might take a target and put it into an animal to immunize it and basically have the animal create antibodies against the target, which we would then extract from the animal, or we might run a phage or a yeast display library, where you might have a fixed library, so a bunch of random sequences that have been fixed that we try to latch onto a target.
What's exciting about AI is actually now, if I want to cut down those timelines and also go after those harder targets, I might take a target, generate those molecules with the AI model, and then go straight into the lab and order DNA for each of these. Whereas before I might have used the same library each of the times, or I might not have even used a library at all, you're now for every target, actually having AI-driven repertoires or AI-driven sequences that we're going straight to testing in the lab.
I can imagine that this is again, another place where we have a close partnership with Twist, where we can very quickly send sequences to Twist on a Monday, for instance, and then a couple of days later, actually have DNA synthesized so that we can go and confirm whether these designs actually work in the lab. To give you an example, a case study of how we might use a method with these success rates in order to either create data sets or validate our models, let's talk about that drug developability challenge. One of the exciting things we're seeing with the latest models is that we can generate an antibody that doesn't just bind the target, like I showed on the last slide, but actually has drug developability properties like thermostability, polyreactivity, self-interaction, hydrophobicity, as well as a host of manufacturing properties.
Now again, let's think about how we actually build this data set. We would take a bunch of targets for this benchmark. We generate antibodies against each of them. Now we have to again, we go to Twist, for instance, and we generate those sequences, and then we turn them into antibodies, and we measure developability properties. One of the great things about working with a company like Twist is that there's actually expertise in all these areas. You could actually run this entire study at Twist, right? Everything from the gene fragment synthesis to the antibody production to the measuring the developability properties, that can all be done over here. One of the things that we've loved about the partnership with Twist is just how responsive Twist is to some of this feedback from a customer like ours.
They have been since the early days of working together, and I think that's really important if you think about where this industry is headed. The workflows that we are running today in the lab look pretty different than they were even just a year ago, right? Because if you're going from testing a system where one in 1,000 of your antibodies bind versus where 100 in 1,000 or 200 in 1,000 bind, it really changes the kind of experiments that you might want to run and the feedback loops that you'd like to run. If it only takes now 24 hours to design an antibody on the computer, the next bottleneck is actually how fast can actually synthesize that molecule and then run these downstream experiments.
We're really happy that Twist is continuing to invest in this and continuing to make those timelines faster, because that's something that can directly benefit the feedback loops through which we can validate our models and then also continue to build their training sets. Maybe lastly, the thing I'll say is that this is a really exciting area to be building in right now. I think both the advancement in the experimental methods of how fast we can turn things around, and then also just the pace at which the models are developing really creates an interesting flywheel where as the models get better, there's more DNA that we need to order, and then as we can synthesize DNA faster and we can run these experiments faster, the models update faster. This whole thing goes into a really nice flywheel.
As a company like Chai, I skipped over this slide earlier. The company's raised about a $250 million of capital in order to go and push the frontier on these models and then take those models and deploy them in some of the largest pharma companies in the world. In order to do all of that that quickly, we've needed to work with partners like Twist who've been able to iterate really quickly on our needs, able to push the bounds on how many of these antibodies we can actually produce and the amount of sequences we can measure. This has been extremely productive for us because it means that we haven't had to go and develop all these things in-house.
We can instead rely on a partner like Twist who has the expertise to really deliver on the scale that we need in order to push forward the bounds of AI. Thank you very much for having us. Hope the rest of your day goes great.
All right. Fantastic. Chai Discovery, they do not have a wet lab, right? You heard from them directly as to how they leverage our services, and they use a wide range of services. We've worked with Joshua for many years in different capacities, and so it was a great example of a very satisfied customer who continues to push the bounds of research. Now, I'm pleased to introduce a real live person in the room, Dillon Flood, who is co-founder and Scientific Director of Elsie Bio, a wholly owned subsidiary of GSK, where he leads the development in next- generation oligonucleotide therapeutics and RNA engineering technologies. Our objective is to show a lot of different customers doing a lot of different things. Dillon, over to you. Thank you so much for being here.
Thank you. Yeah, thanks for having me today, guys. I'm excited to show you a little bit about what we're doing at Elsie Bio. Like Angela mentioned, we were a small, scrappy San Diego biotech company that started in about 2021, and in 2024 became a part of a much larger company called GSK. It's been a wild ride. It's been super fun. I'll tell you today about our really interesting technology that we built that allows us to really increase the efficiency of oligonucleotide screening and selection to make better drugs. We did that with a lot of collaboration with Siyuan and his chemistry team here. As myself as a trained chemist, it's been wonderful to be able to pitch these crazy types of requests to Siyuan and their team and have them come back and say, "You know what? That's wild, but it might work.
Let's go for it. With that, we've opened up some really big doors, and instead of using this technology to train models for antibody type drug discovery, I'll cut to the chase, we're now using it to look at oligonucleotide drug discovery. Kind of a different bend, but it all comes back to DNA writing. What are oligonucleotide therapeutics? These are short synthetic pieces of RNA or DNA that are used to target RNA or DNA in the cell. What we typically think of when we talk about RNA therapeutics are things called antisense oligonucleotides and short interfering RNAs. I'm sure you guys have heard of them. They're fantastic drug modalities. For a long time, folks thought these things were extremely programmable.
When we started the company about five years ago, there was a lot of dogma in the field, and people thought that these things were super programmable. You could use Watson- Crick- Franklin base pairing to program your sequence. You slap on a few patterns, and you're good to go. That might have worked for some of the early targets, but now that these things are going from rare diseases to common diseases, we need better therapeutics. One of the things that makes this hard is therapeutic sequence prediction. You can very easily predict short range sequence shape and folding on shorter RNAs, but once you get to full target length, it gets really hard. That's because these things are not a one-dimensional string of A, C, Gs and Cs, right?
They're a dynamic folded structure that has secondary structure, it has tertiary structure, and a really kind of underdefined protein RNA interactome. What we did is developed a technology that we call ROSALIND. It's a DNA-encoded library based technology that allows us to rapidly screen massive amounts of chemical diversity through our platform. This allows us to perform our selection techniques in a system that natively recapitulates the RNA structure and hopefully that binding interactome as well. Why did we develop ROSALIND? It was really to get at this eliminating what we call the oligo design question for any one target. There's tens of thousands of sequences you can make to that target. There's tens of thousands of patterns of chemical modifications you can pattern on that sequence, and there's tens of thousands of ways to stitch these things together.
I'm not a mathematician, but I'm trusting the people who put that on the slide. There's a lot of ways to put these things together. What we tried to do was come up with a way that we could take a much larger chunk of the pie to really look at the chemical diversity that we're seeing in this space. That's because most folks, if you look at a lot of the drugs on the market now were discovered after focused, dogmatic-type drug discovery efforts where people looked at 100 to maybe 500 different constructs. We think there's just so much more out there to explore that we can find better molecules out there, even in crowded spaces. What we did is we developed our ROSALIND platform. Again, it's a DNA-encoded discovery engine.
That's all super fun and great, what it actually allows us to do is increase the amount of screening and selection we can do at these various stages by orders of magnitude. When we look for our sequence to any target, we're now, instead of screening a couple of 100, we are screening 10 to 100,000 constructs in a single tube all at once. This is how we define our selectivity and our activity of our constructs. We take one of those winner sequences, and we start to pattern on modifications to these things. DNA is a great format for storing information and doing all sorts of stuff, but it's not a really good drug, so we need to modify it. We take our best sequence, and we start patterning in our modification patterns.
We can do this on the scale of 10 to 50,000 constructs at a time. This really helps us fine-tune our toxicity and our PK/PD profile. The last bar yet is not yet DNA encoded, but what we can do is fine-tune those properties through optimizing the linkages that link all these nucleotides together. I'm going to show you some data. It's a little embarrassing because this is all from five years ago. It's the oldest data we have. It's all the lawyers let out the door. Even the first time we ran an oligo discovery campaign with our ROSALIND platform, we were only able to do that because Siyuan picked up my phone call and said, "Sure, I think we can put degenerate bases in there." That was the start of a great partnership.
What we did here is we took a highly studied target transcript. This is for TTR. It was the benchmark case for all the RNA companies out there, Alnylam, Ionis, so on and so forth. We said, "We're going to go find a better sequence to target this gene." What most of these companies have done is they've all landed in a similar locus, which is over by the 3' UTR. What we did is ran our exhaustive screen, our exhaustive search of this transcript, and we found constructs that were distal to that area that were about 20x more potent. We thought that was proof positive and increased screening would give better results. What we did then next is take our single sequence. We threw away the dogmatic gapmer-type approach and started patterning in modified nucleotides into these things.
What we knew about that was that that would help us modulate the protein binding effects of these strands of nucleic acids. What we could see here is that we could keep our high activity constructs while really modulating our toxicity profiles. What we could then do is take that second-generation lead, and we can pattern in with some really cool chemistry that my other co-founder at Elsie developed. We could fine-tune the properties of these constructs. I won't spend too much time on that because I want to get to what I think is even cooler, which is where we are now. Back in the day, we were producing libraries with Twist with all DNA constructs. We were then brute-forcing our Gen 2 constructs with classical column-based synthesis approaches.
Thanks to another phone call that Siyuan graciously picked up, we've now been able to incorporate all sorts of interesting modified sugars into our synthesis process. What this allows us to do is really change the activity and protein binding impact or profiles of these types of molecules. I'll talk about these ones on the right here. It's 2'-MOE and LNA. These are the base state of play for antisense oligonucleotides out there, as well as the PS bond. After much back and forth with Siyuan and his great team, we were able to get some really cool methods that could allow us to incorporate PS bonds into these constructs at high scale. I'm not talking just in a couple of sequences here and there.
This is across the entire chip with complete control, with as good a control as you get with regular DNA. We were then able to incorporate two things, the MOE and LNA, like I mentioned. Really interestingly, we were able to incorporate MOE at very high fidelities. This is notorious for being a hard-to-incorporate base in oligochemistry. The fact that we were able to do it on the silicon chip was amazing to us as chemists, super exciting as a scientist, but really started to enable what we're doing next. What that is building fit-for-purpose data to train our own AI/ ML models to power the entire foundation of our oligo discovery platform now. Right now, we are using this data. Today, we're using this data to start triaging compounds that come out of our selection and screening processes for possible tox effects.
This is great. This is a wonderful use of AI. It takes our screening from hundreds of compounds down to maybe 100, which is a huge win for us in time and money. That said, where I see this going is as we continue to increase this data corpus that we have, we're going to be able to start using this in a generative fashion to pre-design the constructs we want to move forward before we even get to the screening phase. With that, this doesn't happen alone. Our small little team of 20 is now a part of a massive team of about 20,000 researchers. It's been great to integrate into GSK. Very different than the Elsie vibe, but it's fun. With that is the Elsie story, and I'm happy to answer any questions if anybody has any.
Just shout it out?
We'll take this.
Just go?
Matt's got the mic.
Oh
We'll run a mic back to you.
Thanks very much. The predictive toxicology in PK/PD, it seems like that's a huge potential maybe for AI, but the models haven't made a ton of progress yet, and given most failures are the phase II tox phase, I'm just curious, as someone that's worked in the field for a while, what your perspective is on the progress being made on predictive toxicology, and if you think there's potential to improve approval rates over time, if that's an area that AI can contribute to?
Yeah. I think in a therapeutic modality like this, where there are strong class effects, class tox effects, and you can boil that down to interaction with a protein or a class of proteins. There's actually a lot of promise in being able to predict that, right? Because what we're functionally doing here is taking constructs that will fold into three-dimensional shapes, exposing them to these proteins, seeing how they bind those proteins in ultra-high throughput, and using that to make predictions. If you can basically boil down what is causing the tox, and it's something you can address, I think we're going to be able to build the data to actually predict on it, but some of these other larger multi-system tox effects might be harder.
Okay. Can you hear me? No.
Well enough.
Well enough. Try it again? Okay. I'm kind of loud, but you probably want people to hear. You can hear? Okay. All right. Thank you for doing this. In your presentation and in the other presentations, it's been really exciting to hear how quickly AI is leading to development of a higher number of well-profiled drug candidates in much shorter duration. I'm going to actually ask a little bit of a financial modeling question, which I know is a bit unfair, but you're generating a lot of data.
Not my department anymore.
Yeah. You're probably better than me, so don't worry about it. You're generating a lot of data up front. Is there a point where, using you as an example of an important customer for Twist, where you've essentially generated enough data recognizing how much you have, your staffing constraints, your capital constraints, where there's kind of like a bolus of activity with Twist, and then it drops? Or is it the opposite, where there's a consistent build of demand for Twist? I think a lot of us in this room are really excited about this, but I think we're trying to figure out, are customers like you going to spend a lot on Twist up front and then pull back for a while, or does it actually amplify over time?
Yeah. I think that's a really great question, and something that I have to talk to our higher-ups about all the time. What I see is that kind of far-off future where the screening dies off is something that's in the far unforeseeable future just yet. That's when AI takes all of our jobs, and we all don't have to work anymore, right? In the near term, what I see is that we have just started to produce models that are useful and exciting to the internal stakeholders that we're working with. We've been able to deploy these on programs, not only to just reduce tox and triage, but also go the other way and tune in polypharmacologies that are interesting and elusive and hard to get at.
What I see in the next three to five to six to seven years is a steady ramp-up of needing to build data, and this platform relies on ultra-high throughput parallelized synthesis to do that data build. I think there could be a time in the next short to medium term where for a single modality, we have enough data that we can be predictive. As a large pharma, we have a handful of modalities that we want to go after that we can't even start to think about modeling yet because we don't have the data to do it, right? If we're putting our flag in the ground in ASOs and siRNAs, this is a rapidly expanding field. There are self-amplifying RNAs. There are upregulating RNAs. There are RNAs that interact with regRNAs.
This is a whole field of emerging RNA biology that we're still going to go after. ADARs, all these things. I see while I'm still around, I don't see that dying out anytime soon.
Super helpful. Thank you.
Thank you so much. I am continuously inspired by our customers and the endless creativity for the next problem, the harder problem, the deeper they go each and every time, and you just heard it directly from Dillon. It's not a problem that's going to be solved in our lifetimes. With that, we're going to change directions a little bit. We're going to move away from drug discovery, and we're going to focus on enzyme engineering for sustainability. We're really going to change it up. Our next presenter is also joining us via video because she is in Australia. Vanessa Vongsouthi is Research Founder and Head of Bioengineering and Discovery Research for Samsara Eco. She leads development of AI-designed enzymes that enable infinite recycling of plastics and textiles through advanced circular biomanufacturing. Vanessa.
Hello, everyone. Thank you for having me today as part of the Twist Bioscience Investor Day. My name's Vanessa Vongsouthi. I'm one of the research founders and Head of Bioengineering and Discovery Research here at Samsara Eco.
Twist is one of our longest collaborators and enablers in our discovery and scale-up ecosystem here at Samsara. It's a real pleasure to be here today, and I'm really looking forward to taking you through what we're building. At Samsara, our mission is to create circularity for the world's most valuable materials. We started this mission in 2021 by targeting a group of materials that is arguably one of the hardest to ignore, and that's plastic. The world has produced over 10 billion tons of virgin plastic from fossil fuels to date, and this really isn't just a problem of waste. Every kilogram or ton of virgin plastic begins its life as oil or gas, extracted, refined, and then moved through a supply chain that is long, carbon intensive, and increasingly exposed to geopolitical risks.
It's no surprise then, that the production of virgin plastic currently contributes to over 3% of global greenhouse gas emissions, and this is expected to reach 15% by 2050 if we stay on this trajectory. Despite this, only 10% of all the plastics we produce globally get recycled today. This rate is as low as 0.3% when it comes to textile-to-textile recycling. The reality is that no matter how meticulously we sort the plastic waste at home, most of the things that we consume actually don't make the cut for traditional mechanical recycling. Usually, they're too contaminated, contain dyes, or are mixed with other plastics. Textiles have an even slimmer chance of making it through. In practice, it's really only the cleanest, clearest plastics that enter the mechanical recycling waste stream, where they're melted and extruded into recycled plastic.
With each of these cycles, the plastic also loses some of its material quality and strength until it eventually ends up in landfill or has to be blended with increasing amounts of virgin materials. This is really the problem that Samsara was founded to solve. It's what brought us to leveraging enzymes to deliver material circularity. This slide gives you an overview of our technology platform. It's an integrated system that takes end-of-life products and turns them into virgin identical circular materials. At the heart of our platform is a machine learning driven enzyme design engine. The designs we generate are brought to life using Twist's Clonal Genes. They're delivered to our labs, ready to experimentally screen in a 96- well plate format.
Rather than cloning and sequencing genes ourselves, we receive them sequence confirmed and ready to slot into our enzyme screening workflows. This means we can move really quickly from a computational design sequence to an experimental data point, at a pace and scale that just really wouldn't be possible otherwise. From screening our engineered enzymes, we take the most promising candidates forward through characterization and process integration until they ultimately feed into our chemo enzymatic recycling process that you see here on the right. As input to our process, we can take colored, mixed, or degraded plastic from textiles, and then our enzymes get to work breaking them down into their original chemical building blocks, also known as monomers.
These monomers are equivalent to what we have to extract from petrochemicals today, which means we can purify and repolymerize them into virgin quality materials that can be manufactured into brand-new products. Importantly, this really enables infinite recycling, We see no loss in the quality or the integrity of the material, no matter how many cycles it goes through. At the core of everything we do are our enzymes. Specifically, these are new to nature enzymes engineered to break down plastics at speed and scale. These aren't just enzymes borrowed directly from nature and dropped into an industrial process. While some naturally occurring enzymes can degrade plastics, they rarely meet the demands of a real production environment. Often, they're too slow, too unstable, or unable to withstand the operating conditions that we require.
We use natural enzymes only as a starting point and a foundation that we can then build on to create proteins that are optimized for speed, stability, and precision at scale. What makes this challenging but also exciting is the sheer scale of the protein design problem. To give you a sense of this challenge, a typical enzyme with just under 300 amino acids has more possible sequence combinations than there are atoms in the observable universe. The question we face in enzyme design is: how do you search that vast space efficiently to find the very few sequences with the properties you actually need for an industrial process? As you can imagine, searching through by random trial and error would be near impossible. We need smarter, more principled ways to navigate our search.
This is what our platform is built to do. It does turn out that one of the most powerful approaches we can take is to learn not just from the enzymes that exist in nature today, but actually from their entire evolutionary history. Like species, enzymes have a history that stretches back billions of years, and along the way, countless enzyme variants have existed and disappeared. Using a technique known as ancestral sequence reconstruction, we can actually infer all of those earlier sequences and bring them back into view. This matters enormously for our understanding of how different protein families work but also gives us very rich training data.
Rather than only having a few 100 natural sequences that are relevant, leveraging this method gives us tens of thousands of ancestral sequences that allow our deep learning models to identify the patterns that better link things like protein sequence to protein function, activity, and stability. A really great illustration of this is our nylon 6,6 hydrolase, which is estimated to sit at about [10 to the 82nd power] possible sequences away from the closest naturally occurring enzyme. To our knowledge, it was the first ever enzyme that was characterized to be capable of degrading nylon 6,6. This is a breakthrough that was really only reachable because of the richness of this evolutionary data. If we compare mechanical, chemical, and enzymatic recycling approaches side by side, the advantages are clear.
Our process enables a true closed-loop circularity, returning plastics all the way back to their original monomers with no loss of the material quality. It handles the mixed plastics and composites that other technologies reject, and it operates at a low carbon footprint relative to virgin plastic production. Critically, we've demonstrated that the monomers we produce can be repolymerized into virgin identical end products that look, feel, and perform exactly the same way that fossil-derived materials do. Our milestones to date reflect the real-world traction that our technology is generating. On the product side, with our partner, lululemon, we produced the world's first enzymatically recycled nylon 6,6 garment and launched a full retail collection with them. We've also produced a clear recycled PET bottle. These are consumer-facing products that prove our materials are virgin identical in every sense that matters. The industry response has also been equally strong.
Lululemon have committed to sourcing 20% of their fiber portfolio from Samsara over the next 10 years. LSKD, another athleisure brand founded here in Australia, have also followed with a long-term agreement. We've also announced a polyurethane collaboration with The LYCRA Company. Our first facility is now open here in Australia, as you can see in the top right, and our commercial plant is on track for 2028. We've raised $107 million to get there, and so the technology works, the market is ready, and we now have the backing to build this. Plastics are really just the beginning. At Samsara, we believe biology scaled into industrial processes is one of the most powerful tools we have for addressing the material challenges of our time.
Our platform is built on the marriage of protein design, process chemistry, and engineering, and it's that combination that makes what we do differentiated. We built this platform for plastics, but it's designed to go much further. The same approach of co-designing proteins and processes at scale applies equally to other polymers, green chemicals, critical minerals, and even carbon capture. We're not building a single product company, we're building the infrastructure to scale biology into industry. Thank you so much for your time and for the opportunity to present at the Twist Bioscience Investor Day. On a personal note, I've been working in this field for nearly a decade now, and the difference that Twist technology has made to the pace and discovery of innovation is something that we genuinely feel firsthand at Samsara every single day, and it's just fundamentally changed what's possible for companies like ours.
If you'd like to continue the conversation, please don't hesitate to reach out, and you can also follow Samsara Eco on LinkedIn. Thank you.
All right. I'm going to change directions once again. At this point in time, we are going to say goodbye to our friends on the web.