Talking Biotech 445 - Designing New Proteins with A.I. - Stefan Van Grieken

Hi everybody, and welcome to this week's Talking Biotech podcast by Calabria. Now, proteins— they're the shakers and movers of the biochemical world. Proteins are the catalytic and structural elements of cells. They're the enzymes, the ion channels, the constituents of the ribosomes themselves that make more proteins. So proteins are the central players in everything from metabolism, developmental changes, many aspects of physiology. It's across all of life too, from bacteria all the way up to animals are like us. And evolution is what shaped protein function. So proteins are composed of amino acids, which each one has a specific chemical characteristic, and they're put together in a chain. And when that chain is placed into a solvent like the cytosol of a cell, they take on a discrete shape that helps with their function of either binding another partner or substrate or whatever they're there to do. And proteins are around us in places we don't realize, like our laundry detergents, industrial chemistry that degrade cellulose or other types of manufacturing applications. So lots of places where proteins matter. But how do you make proteins better? And evolution did a pretty good job and made proteins pretty much to optimize for their given purpose. But can we make the wheel rounder? And scientists have tried to do this for a long time. You can change that amino acid sequence to see if you make it better or worse. The problem is, is that that's a glacial process if you're trying to engineer large proteins on a global scale or make fast changes. And so can there be a faster way to do it by incorporating machine learning and AI? And that's what our guest does today. We're speaking with Stef van Grijken. He's the CEO and co-founder of Cradle. Welcome to the podcast, Stef.

Thanks for having me.

Yeah, this is great. So I, I, let's start with the idea of Cradle in general. So what's the business model and how are you applying these tools to change proteins in protein space?

Yeah, sure. So Cradle is a software business, and really what we'd love to build is Photoshop for organic chemistry, where any scientists like yourself could go into the lab and be like, ooh, I kind of love maybe this protein I have, but it's a bunch of things it doesn't do, or a bunch of properties that I'm unhappy with. And what you can do at Cradle is akin to how you might generate a picture through a picture model or get a question answered through something like ChatGPT. you can generate a library that you can put into your experiments that hopefully get you a better protein as well as get you to that SAD protein as a marketable product faster.

And I mentioned this kind of in the introduction, but why focus on proteins? Why not just look at DNA sequence and see if we change DNA sequence, could we predict outcomes that way too?

Yeah, I mean, you have to pick a lane. And so before starting the company, I'm a computer scientist by training, so not a biologist. And so before starting the company, I was part of Google Research, where we were doing a lot of work on these large language models that you can now see a lot in places like ChatGPT and other applications. And I bumped into a guy who's now my co-founder, Eli. He's like, I'm applying this stuff to sort of protein sequences. I'm like, why would you want to do that? Because he's like, well, you might not realize it, but they're sort of around you everywhere. Like 60% of all of our drugs are proteins and your detergents, the example you used before, but breaking and brewing and sort of— they're everywhere and they're in all these products. And did you know it takes absolutely forever to develop one of these proteins? And I think sort of akin to how we've learned human language by looking at all the things that humans have written on the internet and putting them into these transformer models, I think you might be able to do something similar with organic chemistry. And so that absolutely blew my mind. And I think as a good tech product person, I started interviewing people. So I interviewed about 100 people that are scientists in laboratories. I actually spent 2 weeks sort of expressing protein and pipetting to learn a little bit more about their space. And I think I really realized sort of 3 things. One, you know, it's really hard to change an almost alien programming language by flipping bits, right? Especially if it takes, you know, weeks to sometimes months to then actually test what the results are of your changes, which makes it, you know, makes it very expensive. Second, I think it's almost like dematrics, right? Protein scientists, they can read these MSAs and sort of look at them and be like, ooh, disulfide bridge, don't like that, right? But it's not a very democratized field yet, right? And so there's a lot of sort of individual specialists there, which makes it harder for a lot of people to develop these programs. And then I think third was more of a technical problem. So I think anything at the cell level, is too complicated. There's too many interactions, and I don't think computers have a reasonable way of learning that in the near term. Also because the ways in which we can introspect them is really hard, right? Like the nice thing with an in vitro assay for a protein is we can sort of do analytical chemistry on them in a sort of relatively isolated way, whereas even a very simple cell is quite complicated. And so proteins felt like sort of an area where tooling could be built, and it turns out that was true.

It's all really cool stuff and I love the idea, but here's the thing that kind of sinks into my mind as a scientist who thinks about evolution and domestication, those kinds of things, is isn't there a certain kind of hubris that says we can do better than a billion years of evolution?

So obviously there's lots of proteins out there that are already great. And I think a lot of people have built products where they take an enzyme from known homology and they just turn it into a product. And similarly for certain therapeutics or diagnostics. It's the same thing. I think very often we want, let's say, things that nature doesn't select for from these molecules. Like, let's take antibody therapeutics. If the antibodies were to automagically get rid of our cancer, we wouldn't have antibody therapeutics, right, in oncology. And so typically, we're trying to do something that's new to nature, or where no clear selection pressure at least exists right now. And that can be in the application, right? So I think antibodies are a good example. But it can also be a property of a molecule, right? Like nature typically selects for things that do pretty well in ambient temperatures, right? But sometimes you want to have a thermophile because it's much more helpful to have something that's much more stable at a higher temperature because you want to use it in some industrial process. Or, I mean, nature selects for things being reasonably fast. If you're trying to develop, I don't know, some detergent and you would like it to work at a lower temperature, you also want it to have a faster catalytic reaction, There was really no reason for nature to have faster detergents, right? And so the humans want different things from the biology, and so we can sort of help them get to those types of product profiles, if you will, or product specifications that may be a little bit outside of the evolutionary manifold.

Yeah, that's an excellent answer. That really opens my mind up on this because, of course, we're oftentimes requiring proteins to do their job outside of their native environments. So why not Use AI to help us design those. That's really cool. Could you take me through the workflow on this? How does it normally work if, does a client come to you and say, I need a detergent that works better at a lower temperature? So like, how does that work?

Yeah, so let me first zoom out to sort of the phases of typical biological R&D, and then we can talk a little bit about sort of how the tool itself works, right? So in terms of phases, there's typically a phase where people try to identify a starting point, right? In therapeutics, they might do an immunization of an animal and find find a set of initial binders. You know, there may be some natural occurring protein as a starting point for a detergent. Like, you know, in our lab, we've worked on a dehalogenase, like, if you want to cleave those bonds, right, that's a decent starting point that you can find in nature. Or, in certain cases, if there's some structural knowledge, some folks even come to us with, you know, like, quote unquote, de novo proteins, or sort of initial binders to maybe a therapeutic target. that can then be brought forward. Now, typically those are, like we just discussed, they're not sort of great, right? And so in therapeutics, you might want to pick a set of leads that are quite diverse because there's a lot of risk that you might not make it to the end. You then want to go into lead optimization because you don't want them to aggregate. They need to express well, like a lot of it needs to get made in a cell. You want them to be thermally stable because it's kind of annoying if you have to ship them around the world in cryofreezed boxes, right? And so there's lots of things, you know, in the case of therapeutics, you don't want them to cause an unwanted immune response. And so a lot of people have a lot of demands from these products that they're building, and they typically, you know, produce assays or they have assays that they use in order to test for these properties. But then a lot of things sort of counteract each other. And so they have to sort of go through this lead optimization or generally enzyme optimization phase until at some point they can go to their manager and be like, I got it, right? And then in the case of pharma, you go file for IND or initial, you know, clinical experiments or in the case of enzymes, you typically go and find yourself some fermenters to go scale up that process. And so how people use it, literally the only thing you need is a starting sequence. And so it could be an enzyme, could be some antibody, or it could be a library of some sort that you care about. And then you need to tell the, like, you need to prompt, quote unquote, the model with what you want, right? So if we were to take an antibody You might be like, I like this epitope, I like this other epitope, I want very strong binders to these epitopes. I really don't like these other epitopes that are very similar, right? So don't bind to those. You know, I'd love this to not aggregate, yada, yada, yada, yada. And so a little bit like Google Maps, you tell us the destination and then you can sort of work with the ML to restrict it. So maybe you're like, I love Qo as an expression system, or—

Qo.

I would love for you to not touch my little linker IP for a thing that I have here, or please, no cysteines on the surface, right? Or whatever sort of restrictions you want to sort of put on it. That's your prompt, if you will. And then you hit generate. And so what happens then? So there's really 2 parts to our models. There's a predictor, and the predictor sometimes has knowledge of a certain problem like stability of molecules, We've collected a lot of data in our lab such that models know a little bit about protein stability, or it may not if it's a very specific assay that you have. And the predictor is like an art critic, right? So it's somewhat good at its job, but it's not always correct, right? But it gets to see a sequence and it gets to judge it and be like, I think this is going to be a good binder, I think this is going to be stable, and it will give it some sort of score. Then we have a generator. The generator is an artist. And it just produces lots of variants, right? And so we can start like 20,000 computers in a data center and just produce lots of different variants and get them judged. And then at the end, we downselect them to a 96-well, a 384-well, or maybe even a large library that you can put into your experiments. And what we've learned from the 32 molecules that are currently active on our software is that speeds up the process between 1.5 up to 12 times. which is significant given how long some of these projects take to develop, whether that's in a therapeutic or a non-therapeutic context.

And so when you're talking about in a therapeutic context, how does that translate into dollars and cents for a company that's doing R&D on improving one of these molecules?

Yeah, I mean, that really depends, right? Unfortunately, that's the answer always in biology. But if you have you know, let's say, a rare disease, right? That's an area where maybe a lot of people wouldn't play right now, and so I think a lot of people there would maybe try something that they wouldn't do before. But, you know, you have therapeutics where it ranges from anywhere from, let's say, $80,000 to do a round up to several millions of dollars if they do, like, cell-based assays, and so, you know, the savings can be quite significant. Especially 'cause you can typically also be in patent longer, which people like, right? 'Cause they get to market their product for a little bit longer than before. But, you know, as a scientist or a computer scientist, sort of n is 32 is too small of an n. So I don't wanna give you one specific number, right?

No, I totally get it. But so you're talking about a training set that the predictor uses, like training data that the predictor parses to identify patterns that can potentially lead to improved molecules at the construction side.

Yep.

It really does rely though on what we already know, or can it kind of make predictions about other enzymes that maybe have never been tested or other proteins have never been tested that just have a patternicity that may be relatable to the function of question?

I think that's a great question. So there's a big difference between, let's say, human language models and protein language models. So for human language, we have a pretty good sampling of human language on the internet. So, you can argue most things have been said and it's pretty complete. The second thing is we as humans can introspect language. We can say if it's correct or incorrect, we can fact-check it if we want. That's obviously not the case for biology. So, how do you do this in human language and how do you do this in biology? In biology, we just have a lot of sequences and nobody knows really how to read them and interpret them. And so for a human language model, the way that you would train it is, at the most fundamental level, is through an infilling task. So you take a sentence and you start removing words, a little bit like how you teach your children human language, right? So something like, Kevin is famous for bleep, right? And then if the model says strawberry or genome, points, right? If the— I know it says butcher, like probably no points. And so if you can, if you do that for every sentence on the internet, you get a sort of what's a reasonable next thing that should go there. And obviously you can do that forever longer sentences. And this is how, if you, you know, I'm oversimplifying, at some point you get ChatGPT. The problem is it doesn't always know what's correct or not. And so then the second trick that you apply is given a certain question, is this a better answer than the other one? We can do something similar for amino acids. So we take some collection of sequences and we're like, if I remove the valine here, which should go here, if the model's like valine, it's like, yay, cysteine, right? And so you get a measure of plausibility at the different positions, and you can do that over longer distances as well. And so that kind of constrains your space a little bit.

Yeah.

However, you don't know, like, will putting a valine here help with binding affinity, right? Or will it help with improving stability of this protein? And this is really one of the big problems in biology. We have very little data. We have very few sequences, right? Evolution only sampled a very limited amount of things that are plausible. We have only done NGS on a very few— small subset of that. And then we have only very few labels, right? Like a lot of things haven't been tested for anything in a laboratory. And so if you, you know, we're in Switzerland, sort of the joke we sometimes make, if proteins were a Swiss cheese, it's mostly a hole, right? And there's very little cheese, right? And so what you need to do from the sequence level is you need to be able to translate that to things humans want, like a stable protein. And so one of the things we've done is we've collected like 40,000 different melting temperatures of proteins. And so models start to learn, ooh, like this kind of little pattern is probably bad, right? So I should avoid it when I'm generating that pattern. And so it learns sort of master-student level understanding of what a stable protein looks like. And so these models are pretty good from the get-go, but they're not experts and they might go out of domain given what you're trying to do. And then sort of the last thing that you can do is you can fine-tune an active learn, if you will, on a set of very small batch, which is, you know, a microtiter plate of your nanoDSF in order to establish, let's say, the stability of these things. So that's how they work, sort of the gut of them. Obviously there's a lot more complexity there as well, but that's the general analogy to sort of human language.

We're speaking with Stef van Grijken. He's the CEO and founder of Cradle. This is the Talking Biotech Podcast by Collabora, and we'll be back in just a moment. And now we're back on the Talking Biotech Podcast. guest by Collabora, and we're speaking with Stef van Grijken. He's the CEO and co-founder of Cradle, and we're talking about using AI and machine learning to make predictions about proteins to improve their function for different applications. And so when you look across the portfolio of things you've already done and companies you've worked with, how do you break down the amount that has been used for therapeutics versus, say, industrial enzymes or maybe other potential classes of, of molecules you would design?

Yeah, I think that's a great question. So in terms of our business strategy, we did decide to initially focus more on biopharma. The reason there is they're typically earlier adopters, the stakes are a lot higher, right? And you need to build a viable business as well. It's a little bit like sort of Tesla in the sense that you first built the car for the richest people and you sort of slowly work your way down and make them more available and affordable for others. So right now I would say it's around 70/30, so 70% therapeutics, 30% non-therapeutic applications.

Yeah, very good. And so what are the kind of the next realm of applications that you might be taking on? I mean, think about the way I think about the way that different enzymes are used in either food processing or maybe even in fermentation or whatever. And what's the next big domain?

So, I think in therapeutics, we cover a lot of the major modalities. So, antibodies, recombinant proteins, vaccines, those types of things, but there are more, let's say, smaller modalities that are currently coming up. So, for example, in peptides, I'd love to do more like non-natural amino acids and those types of things. So, you could be even more creative in the types of approaches, CAR proteins, those types of things in that space. I think for enzymes, we're pretty good. We have pretty decent coverage of most things that we've seen, like we've seen transporters, membrane-bound things, and so I'm pretty happy in that space. I think in terms of capabilities, so I think an open problem in biology is sort of develop something that has a novel catalytic reaction that doesn't exist in nature, because we still very much depend on sort of known homology for starting points. That's really a frontier. I'd love to have a world where you could have 2 SMILE strings and be like, I want to go from this to this, and figure out a structure that would work for that. I think secondly, we're seeing some early sprouts in terms of finding de novo binders against targets in the therapeutic space, but that's really very hit or miss still. So, getting that to work reliably could be very interesting for diseases that are hard to target with immunization or some of the more common techniques that we have to find initial hits. So those are 2 areas I'm super excited about, because it would sort of open it up a lot more. I think on the platform side, we really want to democratize access to a lot of this stuff. And so making this available to ever more awesome use cases. As a computer scientist, I guess for y'all, this is all pretty everyday, but I see just the sheer amounts of cool applications that you could be building with biology. are just absolutely mind-boggling. And so, I just love— we're sort of app stores, and you have yet another video or social app, meh. I'd love to sort of cure more disease, get more things to not use hydrocarbons or animals, and sort of build more cool stuff with biology. In my career, I've always been a plumber for other people. And I worked on Android, which is an operating system for mobile phones. I worked on TensorFlow, which is the software people sometimes use to build machine learning models. And so, you know, hopefully I can make a positive contribution with this company to more people, you know, like yourself doing a lot more cool stuff with bio.

Well, what about interesting issues for the planet? I mean, there's a whole host of things like maybe enzymes that can degrade an oil spill better or, you know, chew up polyethylene plastic or, you know, degrade the bottles that are all over the ocean. How likely is it that this kind of technology can be used for applications in helping us remediate the residues of human populations?

Yeah, for sure. It's actually one of the reasons I got super excited about the space, right? And it's anywhere from, you know, one of my co-founders, Elise, she was at Perfect Day Foods, where they're making whey protein without animals, right? I think to your point, you know, plastic degradation, you know, I think sort of zooming out, One of the cool things about biology is this evolution thing seems to have tried to find the lowest energy state to do very complicated catalysis. Ultimately, biology will win. We just need to find the right sequences. And then anything where we currently use hydrocarbons, or we use animals, or we use plants in a way that we shouldn't be using them to make the things we want, while being able to produce them with sort of your local sugar, water, and carbon source, with no waste, with no carcinogenics, with no whatever. That's sort of the future I want to enable. I think what stands in the way of that is you need to pony up like $100 million and a lot of insurance to develop. I always tell my computer science friends, God left the hard problems for the biologists, right? Yeah. So I think to your point, it's just the economics of a lot of these things don't make sense today. But let's say if we could reduce the time it takes to develop these things to, let's say, 2 to 3 years, and you would need maybe $10 to $20 million, I think we're going to see a lot more things built with biology.

Oh, very good. And so, but what's real and what's not? I mean, where are the limitations of this technology? Because we hear everybody talk about AI and designing the next generation of molecules. You know, where are the limitations?

Yeah, for sure. I think I would say we're still in sort of GPT-0.2 when it comes to protein language models. So there's a few major limitations. One is we still can't reliably produce things that are very far away from something nature has already done, right? And so in an antibody, we can change, let's say, a CDR region to have more affinity to something, but producing completely novel binders to all the targets in the human protein repertoire reliably doesn't work. We have some initial things that work, but they're still— we're still not there. I think there's still lots of open problems. For example, in the case of therapeutic, like immunogenicity, right? Sort of, are you gonna trigger an unwanted immune response by introducing this molecule? It's hard to say, right? Immune system is a very complicated thing, and how it responds to a foreign or a protein that's non-natural sort of is still very hit or miss. And it's a major problem in the industry, right? 'Cause a lot of things do fail in clinic 'cause they produce too many anti-drug antibodies, and you ultimately have to shut your program down. I think the ways in which these models can introspect more complicated biology is still very limited. So sort of cell-based assays you can still learn from, But beyond that, sort of small animal studies or even sort of more complicated in vitro systems is still an open question. We talked a little bit about metabolic engineering, sort of anything that's sort of at the level of a full organism. And, you know, I mean, you work on plants, right? And we were joking at the start of the recording, yeah, at least the bacteria grow a bit faster, right? And we get to learn a lot faster, right? But sort of the more complicated the system, the more the harder it is. And then, you know, a lot of the, quote unquote, de novo stuff that we see for our space is still very much about finding the starting points. But we still haven't found models that only produce stable variants or only produce variants that express well in any expression system of your choice. And so we're starting to get to places. You can't prompt these models very well still, like you can't really give human preference to them as much. Um, so I would say, you know, um, uh, at least what we built at Cradle is still very much the minimum viable product. It's starting to have impact, uh, but, you know, there's a long way to go to, to make this even more broadly applicable and solve more problems for people.

So who is your client? And like, so if somebody comes to you, uh, with a question and they say, we would like you to try to design blank for me, are they typically pharma companies, or who's, who's coming to you for doing this, and how would they interact with you?

Yeah, sure. So right now, our focus is sort of global players in pharma, materials, food, ag, and so we typically work with the larger customers. The teams we work with, they're the scientists that, you know, are within the individual therapeutic teams or enzyme teams that are developing the specific project. So, you know, preclinical R&D in the case of pharma, and typically just R&D in the case of food and chemicals and ag. The way they use it, it's a web application. So you go to app.cradle.bio. You sign in with your credentials. You tell us something about what you want. And then literally a library comes out, and you call Twist, and you get some DNA back. And you go try it in your lab. So it's that simple. We recently won the Adaptive competition. Adaptive is a CRO that does a lot of antibody work. where about 130 teams participated to build a better binder against EGFR, which is epidermal growth factor receptor. And there is a commercial product in the market called cetuximab, marketed by Merck, and I think it's outlicensed to a couple of people. But, you know, that was— that molecule had a 9-point-something binding affinity, a nanomole binding affinity to the target. We had an intern use Cradle for 4 hours, and they produced you know, a variant that had a 1.1 nanomole binding affinity to the target. And so, you know, intern in 4 hours. So I think we're getting into a fun space, right, where that is possible. Now, obviously, a binder doesn't make a drug, right? So I do want to call that out. But, you know, it is a helpful starting point if you want something that's efficacious.

Well, at the very least, if it's not a drug, it's another data point and can help you build towards that drug, which is really, you know—

Certainly.

—the process, right? The more you're learning from this, the more you're able to build into it, which is really kind of a nice aspect of this type of process. But what about issues of safety? So you're redesigning proteins that are being used as therapeutics, and how are you certain that they're binding something that is discrete and maybe not having collateral effects?

Yeah, for sure. So there's a couple of ways in which we can do this. I think for toxicity, I think models have— there are existing models even out there that are pretty good at predicting that somewhat. But ultimately, the proof is obviously in the experiment. I think like any other assay, we can optimize for it. So if you were to run some tox assay or immunogenicity, obviously for biologics is another major issue. If you were to run like an HLA binding assay or a T-cell activation assay, like those types of things, can sort of do. But, you know, it's still a major issue for the industry. And, you know, the immune system is a complicated thing to fully introspect. I do think we are starting to see some signs of certain things you can learn without a lot of laboratory data. So, for example, we've seen we can learn expression systems between species. We can learn expression systems Or sorry, we can learn expression between, from one family to the other in the same expression system. Hopefully at some point we will get better at eliminating any form of liability in the therapeutic that might trigger an unwanted immune or toxicity response. But that is still somewhat out there in the future.

You mentioned that an intern can use your software to develop a molecule with a higher binding affinity that can interact with biology. But can't this same kind of approach be used in nefarious ways, like to design molecules that can be deleterious to biology, maybe to, you know, even target as a biological weapon?

I think that's a really great question. And especially with a lot of these models being able to go much further away, right? I sometimes use the analogy of the Old English went to sea and they had the year be dragons. We now have speedboats, right? And we can go places that are a lot further away from what you might typically do. There's a couple of things that we can do in order to try and take this obviously very seriously. So, on the one hand, we obviously have our terms of service. We explicitly forbid certain application types, like we don't want people to make weapons and those types of things. There are certain biological systems we don't want people to interfere with, so cloning propagation, We don't want people to encode for a known toxin, a known carcinogenic, those types of things. And then thirdly, there's a couple of, you know, and honestly, coming from tech, I find biology has really done this quite well. There's a lot of industry associations that sort of protect. And so, you know, obviously, when proteins get generated, we can look for patterns of nefarious behavior. You know, if you order your libraries from a DNA provider, they will do that as well. So, that's something that we can do. But I think it's an important point. Luckily, a lot of the companies that are active in this space, including a lot of folks in the research community, we've come together in sort of a coalition to actually think about these ideas, share thoughts, and sort of make sure that we can keep this stuff safe now and in the future. I think for us specifically, given that we work with some of the largest companies in the world, right? Novo Nordisk is a company of ours— sorry, a client of ours. Johnson Johnson, companies like NovoNisis or IFF in the non-pharma space. We do trust them to not be out there for something nefarious. Once we start opening up the access to these capabilities to smaller companies or maybe even individuals at some point, I think you do need to get even a lot more serious. I think finally, you know, obviously your audience is mostly biologists, but obviously proteins don't self-replicate. But, you know, once we get into a world where it's easy to program cellular level, it gets even more tricky, right? But that's, I think, probably still several decades out, or at least I don't want to speculate on that too much.

Well, if you did want to speculate on something and had a crystal ball in front of you, what is a problem that you feel is close to being cracked using this technology, either from knowledge you have from something you've already done, or maybe something that you feel the industry is really focusing on? Like, what's the next big breakthrough we might expect from AI-enhanced interrogation of proteins?

Yeah, it's a very good question. I mean, it's always hard to say. I was at Google Research and everybody's like, oh, nothing's gonna happen for the next few years. And then boom, right? There was something new and you were like, oh my God, right? This is cool. I actually think a lot of it is maybe going to be more at the level of the industry than anything else. So there's a few trends that I'm seeing that I'm excited about. I think the first trend for me is we are seeing between 1.2 up to 12 times speedups in some of these programs, and we're still early in the ML. So let's say now you can build a bio-based product for maybe $10 or $20 million with a small group of people, a little bit more like how app development became a thing, right? where before you needed mainframes in your basement and all kinds of other stuff. Let's say the access to developing a biobased product becomes less risky and cheaper. I think that's going to help a lot more people that are hopefully listening today build new companies. And so that's one. So if the success rates go up and the speed to get to market goes up, I think things change a little bit. Because especially if you look at a lot of these biobased products today, you almost need to be a bank in order to bankroll the risk. Sort of, you need, I think, of every 25 projects that start in therapeutics, only 1 makes it through to the FDA. And so that's a success rate that's lower than venture capital, right, in many cases. And so what I hope will happen, and you have some other tailwinds here as well, like even throughout the period that we've been building Cradle, which is the last 3 years, ordering a library from Twist, went from $25 per sequence to now maybe a couple of dollars per sequence, right? And so you see orders of magnitude of improvement. Like if you look at analytical chemistry applications like SPR and those types of techniques, they're becoming more readily available to people as well. So I think the innovation is not going to be like one ML trick. It's not going to be one software provider like us. I think it's going to be a combination of a lot of things that are ultimately going to make a lot more people build stuff with biology. And I just think, you know, I was joking, I think antibody therapeutics is roughly the same revenue as online advertising, right? I joke, I was at Google before, and so I joke to my former colleagues about this. It's just the output markets in biology are just vast, right? And I think a lot of people are going to be super creative. And I tell my daughter she should become a biologist, right?

I tell my daughter she shouldn't.

You've had to live through it, right? I'm just doing math, right?

Well, Stefan Greichen, if people want to know more about Cradle and maybe sign up for use of the services, where do they do that?

Yeah, sure. You can go to cradle.bio and then hit early access, or you can send me an email at stef@cradle.bio or follow me on LinkedIn. Stefan Grieken. I'm, I'm, uh, I'm gonna be okay with this. G-R-I-E-K-E-N. Weird Dutch name, my apologies. Uh, most, uh, Grieken, most, most peoples cannot do the sort of sound, right?

Well, I don't try to do it because I don't like to appropriate another language's, uh, gurgle noise.

So I think, I think, uh, I think it was more, more to get the spelling correct.

Yeah, so yeah, but it's G-R-I-E-K-E-N.

Yeah, G-R-I-E-K-E-N.

Yeah, I took German in high school, so I kind of get the I-N-E thing a little bit. So I was hoping to get that right. So Stefan Grieken. Okay, well, thank you very much.

There you go.

Well, thank you very much for joining me today, and best wishes. And as things improve and you see some exciting outputs, please join us again, okay?

We will, absolutely. Thank you, Kevin.

And this is a really exciting way of merging computer science and biology. How can the, the universe of known protein sequences inform the development of new ones? And this is so exciting because 5 years ago, I don't think that we would have thought that this would have been possible. I mean, maybe would have had dreams that it could have been done, but not an actual pipeline to get it done. Really exciting to see these things starting to happen. So tell a friend, write a review on iTunes or the place that you listen to podcast media. Let's continue to grow the Talking Biotech audience. This is the Talking Biotechs podcast by Collabra, and we'll talk to you again next week.