172. Generating Needles in Haystacks: Elise de Reus Designs Proteins with Cradle

Elise So when you're engineering proteins, you are trying to find this needle in a haystack.

Elise and there's maybe, a couple of needles in that haystack, but you are looking for the solution that meets all these different requirements.

Karl Hey Ron, what's going on?

Erum feeling a little sad. I said goodbye to chat GPT after three years.

Karl Wait, why did you do that?

Erum Well, for the same reasons that a lot of people have done that. Um, for one, questionable antics of OpenAI and Sam Altman is one thing that I was like, Hmm, not really sure if I wanna be providing, My time and my dollars and my intelligence into this, platform in particular. also the whole resist and unsubscribe movement I think is really interesting for multiple reasons. not necessarily like. Trying to like stick it to the man or whatever, as much as like, Hey, we can collectively do some damage [00:01:00] in a nonviolent way And then also, you know, I hear great things about, Claude and it reached a lot of benchmarks. I have friends that are in the space and they're like, it's actually very good. it has been doing, very well. And I was like, you know what? It's time. But it was sad. It was sad because I did say goodbye.

Erum We did have a, a back and forth,

Karl I've used both Chat, g, PT, and Claude, but Claude has been my go-to for a long time. I do use the other models.

Karl I've got, minstrel, the French model on my phone. I use Kimmy. I've used Deep Seek, I've used Grok even. They all have their own uses, but Claude has been my go-to, and I too have unsubscribed from Chatt. so. Sad, a little sad to not use it as much. but Atropic was in the news because they said no to the Pentagon.

Karl They didn't want their models used for, to drive, I'll say weaponry or surveillance of the American public, which is notable. And, I noticed that there was a lot of all of outages over the weekend because so many people were [00:02:00] migrating to. Claude, so good on them. And congratulations and welcome to the club.

Erum Yeah, yeah, it was good. And one of the things before I left chat, I was like, well, which AI should I go to? Just curious to see what I would say. And it was really interesting because, after three years of use professionally, and I'm, and it knows what I'm working on, and it was like. Well, Claude for strategy, Gemini for operations and perplexity for research, and I was like, okay, that's cool.

Erum I just, and then like use it as. Those workers, right. Don't just use one for everything. Each one has a purpose. And like, look at each one as a team member. And I was like, okay, that makes sense. I'll, do that.

Karl I do use Gemini a lot and, I think we've been using it more at messaging lab. I'm always surprised by how fast it is, but when I've done research projects, perplexity, of course, they, I just think that they're unmatched from a research point of view, but.

Karl I've done projects where I ask the [00:03:00] ai, what to query at different LLMs, and it's always surprising, it's exactly what you said. It's like use. Claude for this, use deep seek for that. Use rock for this. And you get very different kind of work styles with each one. And I think it just makes this moment very interesting and it gives us a lot of opportunities to just, work more deeply.

Karl I'm not gonna say work faster because I don't like saying that, but I think more deeply and efficiently.

Erum It did not propose roc, I think it, like it's mentioned like this is not for you. Like, it was like here are all different ais and here's how you can use them based on what all your work and how you think.

Erum And it was like, really not for you. And I was like, okay.

Karl On the Grow Everything podcast, we focus on the impact of biotechnology across industries. And I wanted to highlight one piece of news that is not about ai, but is about lab grown meat. did you see this story iram?

Erum I've been hearing rumblings, but what's going on particular?

Karl So the cellular ag space has seen [00:04:00] a lot of. Slowness or trouble, or it hasn't been in the news the way it was like three or four years ago where there was a lot of money pouring in and a lot of that money poured in. And like with other industrial biotech sectors, movement tends to be slow. And so, the light stop shining on the progress that people are making.

Karl But there was news last week that said that, Researchers have found a way to convert yeast that is used in brewing into an edible scaffolding material, which is one of those things that really. Takes up a lot of the cost when it comes to cell grown meat. So this breakthrough, which has been kind of quietly happening in the background along with others, just kind of reflects the momentum in the field.

Karl Companies and research teams are continuing to move forward. They're optimizing bioreactor design. We've talked a lot about the bioreactors here on Grow Everything. They're optimizing growth media. And even, regulatory pathways are changing to push cultivated meat from these pilot projects to really lead to these bigger [00:05:00] industrial scale productions.

Karl So I'm very excited about this because there hasn't been a lot of lab grown meat news lately.

Erum Yeah. Yeah. So I'm looking at this article and it sounds like this is not necessarily like a commercial breakthrough, but more of a, research breakthrough, from the university. College London, which is great.

Erum And I'm, I'd be curious to learn which companies, you know, work with them or, you know, apply the research into ensuring that, we can get more affordable, tasty, good quality. Lab grown meat, and of course you can't say lab grown. So on the messaging side, cultivated meat sounds much better, more palatable.

Erum but you know, we've been waiting for this and hopefully, you we start seeing it on the shelves or in the. Refrigerated aisle, I guess, maybe on the shelves. If it's shelf stable, who knows?

Karl and one of the things that the article says is the demand for protein continues to increase [00:06:00] as populations around the world grow wealthier.

Karl They want to eat more meat. That requires more. Livestock agriculture and doing, cell grown meat, being able to produce proteins more efficiently Makes a lot of sense.

Karl So, in the past we've mentioned the Coalition for Advanced Biotechnology for Sustainability, and I'm super happy to say that we've worked on a new report with this coalition that focuses on four molecule classes that can unlock and accelerate the bioeconomy.

Karl by the time this episode comes out, the report will be available, so we'll put it in the show notes. We're very excited about that because it focuses on four molecule classes. Terpenes peptides, non catalytic proteins and hydroxy acids. Each of 'em has its place in accelerating the growth of the bioeconomy.

Karl But today we're gonna talk about proteins. So AB four s is a coalition of industry leading or organizations including L'Oreal, ionic, [00:07:00] Lamont, I don't wanna forget anybody, like that's a thing. but also, you know, startups like Cradle Bio and arsenal Bio Yards, who has been on this podcast.

Karl And in their first report, they answered the question, how big is the economic and environmental impact of advanced biotechnology? That the goal of that report was to reframe how industry thinks about deploying biotech and creating biotech. This new report. answers the question of where do we start?

Erum That's really great to hear about the AB four s report. We'll share more once that is released. But today's episode is all about proteins, and we love talking about proteins because they are the workhorses of biology. They do basically everything in living systems from fighting disease in our bodies to breaking down stains in your laundry detergent.

Erum To producing food ingredients and medicine. So everything is based on proteins.

Karl Proteins have [00:08:00] evolved or can be designed or engineered to do very specific tasks. So if you think of a certain medicine, there's a lot of protein-based medicines. Antibodies are produced in your body to fight disease, but they can also be produced in the lab and then injected to fight a disease for you.

Karl And companies that develop things like antibody therapeutics or enzyme for detergents typically runs. Dozens or hundreds of experiments to improve the efficacy, safety, manufacturability of proteins and those runs, or those experimental rounds can cost tens of thousands to hundreds of thousand dollars each time.

Karl So protein engineering traditionally re relies on a lot of trial and error, and it's. Success rate is very low. But our guest today Cradle uses a generative AI model to essentially reverse engineer proteins, letting scientists specify what they want a protein to do, and then receive the possible sequences to [00:09:00] test how would you characterize what Cradle does?

Erum But yeah, you think of it just like as you, work with any ai you're prompting. So it's prompting, but for biology, and the insight is that there are these long sequences of amino acids and making up proteins, and it's like.

Erum An alien programming language, but AI can learn to speak it even though humans can't. So Cradle was founded in 2022. They raised over a hundred million dollars. Their clients include j and j, Novo Nordisk, noesis, and they have really good outcomes on the work.

Erum I mean, there are accelerating all of these studies that have been doing so reducing the trial and error of traditional protein engineering. But let's just stop there. We'll let Elise Deus of Cradle Bio, tell us more about what they're working on.

Erum Elise [00:10:00] Deus, welcome to the Grow Everything podcast. We're so excited to have you.

Elise Thanks Ra. and Carl, I'm really excited to take part today.

Erum Awesome. Okay, well, you know, we love to kick off the episode with learning a little bit about you. So we do know about you. Let's just kick it off. So your journey.

Erum Moves from fungal synthetic biology to high throughput strain engineering. And now you've co-founded an AI driven protein design company. So what was a catalyst that made you shift from doing experiments yourself to building tools that help others design biology?

Elise Well, from an early, early days in my career, I've been driven by impact.

Elise I've been impatient in seeing this impact and that has led me from fungi and studying fungi to realizing the power of being able to use them for producing useful molecules, small molecules or enzymes, to going, well, where can I engineer these organisms more [00:11:00] effectively? which led me into lab automation and robotics, at Zymergen in California among other places, and then gradually realizing that. There are problems that you can address by throwing more robots at the problem instead of me personally doing it myself. But then challenges remain. And so then the question is, can you use AI to improve the quality of the, hypotheses that you're putting into this, experiments? And if I can leverage ai, how can I scale impact even further by?

Elise Trying to get into a platform that scientists across different industries, across different countries, across different organizations can use to effectively design better proteins in this case, in less amount of time. So it's this journey from, doing experiments myself to having robots do them for me, to having ai, running experiments and then creating a platform for others to do the same.

Karl That's so great. But before we go any deeper, can you help our listeners visualize what cradle looks [00:12:00] like? If an R and D team started using your platform, what changes in their work on a day-to-day basis?

Elise So think of Cradle as an AI powered protein design studio. So a scientist has a protein that they're working on, whether this is an enzyme for an a bio, catalysis reaction or it's therapeutic space, an antibody, they open up their cradle workspace.

Elise They outline the project. So what is the protein and what are the properties that this protein needs to hit in order to be commercially viable and safe, and efficacious? then they can upload any data coming in from the wet lab. they can train generative, AI models, to understand the space around.

Elise Like the evolutionary context around their particular project, and then use those models to generate variants of these proteins that are improved according to the properties that the scientist cares [00:13:00] about. Then the output of this in silico design process is file of, protein candidates that can be evaluated in the wet lab.

Elise And then once a scientist gets their results from the wet lab, these can be fed back into the cradle platform to, Very easily and access train models to update their understanding of the sequence to function relationship and, generate that next round, for testing in the wet lab. And so either through the user interface or through APIs, cradles a, an easy way to build custom models for protein engineering projects, ultimately delivering molecules that will work in the real world.

Erum That's really helpful. I could imagine in the lab, even being a former scientist doing high throughput screening, I can certainly relate to all of those points where things can go wrong. And I know I've, clearly, you've experienced this, you know, you were at Zymergen and Perfect Day as well, and you worked on course high throughput microbial engineering.

Erum And in those environments, [00:14:00] they are systems that have huge iteration cycles. Then of course, plenty of points of failure. Was there a breakthrough moment that reshaped how you think about designing proteins or strains?

Elise I think the exploration that led to founding Cradle started with investigating what type of data are we creating in the field of metabolic engineering. And when you look at that, you very quickly realize that the data, we describe it as short and fat. Now that is not necessarily a bad thing, but for machine learning models, it's not the most compatible type of data.

Elise So by shortened that, I mean few data points with a giant number of labels and metadata, a giant number of ways in which those data points are described. and these data points are generated in cycles that take a long time. so if data comes back every couple of months, the speed at which you're learning or at which a model is learning is going to be limited.

Elise And that is combined with the challenge that in [00:15:00] bioengineering we're often taking biological systems out of their natural context into this world in which we subjected to all these different requirements because, enzymes have to all of a sudden work in a different host or in giant industrial fermentation fats.

Elise and that means that we have to require these natural or engineer, these natural systems to fit our. Our needs as, scientists and, bioengineers and putting all of those things together. The question then became, well, where can we produce data in a way that really feeds. Machine learning models and helps us explore more effectively and take bigger jumps to adjusting these natural starting points, if you will, to things that will, be commercially viable.

Elise And for us, that meant scoping it down from, using machine learning on entire sales or entire organisms and scoping it down to starting with. Just quote unquote [00:16:00] proteins. Now, proteins are still a giant, have a, a giant parameter space. I mean, there's more possible configurations of even a small protein than there are atoms in the universe.

Elise But it is a design space that is very tractable with AI to guide the exploration.

Karl Just, so just to give our non-scientific listeners, some background before I get into the next question. Could you, explain why proteins, I mean, you described this gigantic space and how there's so many iterations that are possible, but why is that space so important?

Elise So proteins show up. In all parts of our life.

Elise they are, of course, people know them, as food ingredients. So you can think about alternative proteins, but they're also extremely efficient catalysts. So if you want to convert a, starting like a little building block, like you wanna convert a sugar into a high value, molecule, uh, then. There are enzymes that you can use to catalyze those reactions and, and do that in a very specific way so that you, have natural [00:17:00] conditions, like no harsh conditions, fewer, off products.

Elise So enzymes are a big thing. They even show up in your laundry, detergents to make sure that you can wash at lower temperatures. and then there's all sorts of protein applications in therapeutics. You can think of antibodies that need a bind to cancerous cells. you can think of peptides, now in, in the weight loss, uh, ozempic scene.

Elise and those are just a couple of examples of, where proteins show up in our day-to-day life.

Karl Great. And protein engineering traditionally has been slow, expensive, and super specialized. What's the biggest bottlenecks that you guys at Cradle are trying to remove?

Elise So when you're engineering proteins, you are trying to find this needle in a haystack.

Elise and there's maybe, a couple of needles in that haystack, but you are looking for the solution that meets all these different requirements. And so, what that means in terms of proteins is maybe your protein has to be, stable. So it can't fall apart at a low temperature. it has to sustain higher [00:18:00] temperatures.

Elise it needs to be very specific in the function that it performs. So it's an enzyme. It needs to very specifically turn the substrates into that one product and not, Alternative, or like side waste products. if it's a, an antibody, it needs to bind very specifically to your cancer cells and not your healthy cells.

Elise what ML is extremely, or AI is extremely effective at is, is identifying like finding. The needles that are in haystack or generating needles that are actually better, uh, better suited to the, the, final protein engineering objective. And so, it is in identifying, like in screening hits, finding the right starting points, or generating the right starting points, and then optimizing these across all the different parameters of interest to make sure that you get the, that final protein that, checks all the boxes and.

Elise there. There's a lot of models out there, as you can imagine. But what we're doing with Cradle is, is resolving this bottleneck, or this concern that scientists have, like [00:19:00] with models dropping on a weekly basis. How do I know which model to use to generate sequences? 'cause today you can find in the academic literature.

Elise Models that can generate new protein sequences, but how do you know that they're gonna work for your specific problem? that is maybe not just green fluorescent protein or, some model problem that, was published about. what we're building at Cradle is a system that makes it easy for scientists to use these generative methods.

Elise And ensure that they are benchmarked across very relevant types of protein engineering projects so that our scientific users know when these models are going to work and understand when the models are maybe less sure of an answer so that they understand the confidence of a model, and that they are also using models that have been built and designed to deal with the realities of wet lab data, which is collected.

Elise Across all sorts of different types of workflows, which deals with [00:20:00] things like batch effects, like maybe between one day and the next a buffer was changed, or it's run on a different machine. and that affects, the data that you're feeding into the models. And normal models have to learn the right things.

Elise So all of the, the, the finesse of. Making a model work in a realistic application is what we're solving with Cradle and the product that we're building for other scientists to use.

Erum Yeah, that's really great and I, I love that fact that you were talking about, rather than finding a needle in a haystack, you just generate the needle.

Erum That's so good. It's so good. Multiple needles. Yeah.

Elise Yeah. And maybe you don't wanna put all your eggs in one basket. Like you want different approaches and there's maybe, different, routes to that final solution and, That can also even strengthen an IP portfolio if you have multiple shots on goal.

Elise And they're different shots on goal.

Erum Yeah, exactly. Exactly. And um, and it's also wonderful that, there's people sitting behind the screen making the models that scrape, but we really [00:21:00] need to be grounded in reality. Right. but there's this tension between computational predictions and biological reality.

Erum Can you give us an example of where the model was confident but biology wasn't? And what did you learn from that?

Elise One thing that we notice when scientists want to use machine learning in their protein engineering campaigns is that the quality of data is still extremely important.

Elise So machine learning is not a magic wand that will make all your data quality issues disappear. Our models are tasked with ranking variants from best to worst. So they will try and line them all up according to most active, to least active. For example, now you imagine if, you have a hundred data points from, ranging from, activity readouts from very good candidates to very poor candidates, and each of those data points has a huge error bar, then there's many different ways in which a model could rank those a hundred sequences.

Elise Correctly, quote [00:22:00] unquote. but it's, it's really not, learning the true, ranking. and it's, uh, as a result of, the data, like the noise in the data.

Elise machine learning's not a magic wand to correct for noise in experimental data. The quality of the data points that you feed into these models is still really important. And one place where that shows up is that our models are trained. To rank sequences or protein designs from best to worst.

Elise So from most active to least active. if you're talking about an enzyme, and you can imagine that if there's large error bars for those data points, there's many different ways in which a model can go Well, this ranking is correct, and this other ranking is also correct, and this ranking is also correct.

Elise It's not nothing one-to-one to that, ground truth. Um, and so what's really important is data quality. It's really important to get a sense of how confident are the models that a newly proposed design is actually going to be good. and providing these types of insights to the scientists so that they can decide how [00:23:00] to,

Elise Shape that next experimental round. One place on the cradle platform. Where we offer these insights to scientists is after they've generated a set of, new protein variants based on, models, trained on their data, we have headroom predictions so the models can calculate the probability that these newly designed proteins are going to improve the main, objective by a certain.

Elise amount, or that they're going to meet all the other design constraints. And these signals are, we see that our, scientific users find extremely relevant because it helps to inform how long are we going to push on this project? When is it time to maybe stop?

Karl so one thing is like you mentioned headroom prediction.

Karl Could you just clarify that for our, our listeners?

Elise Yeah, so when you're starting in an r and d project or even halfway through, you don't really know the ground truth of when your project is done. When there is no more room for improvement, sometimes you're running up against [00:24:00] physics or biology and the physics has no, like, this enzyme cannot be any more active without degrading these other properties that you care about.

Elise What scientists in, especially industrial biotech applications are considering is like, how far can we push this and, when should we call it quits? And accept that the designs that are coming out of the experiments are the final designs and that there is no. Further room for improvement and rather than spending an additional couple of months on that next batch of sequences, it's really helpful to have a model.

Elise Say I've been running compute for the past two days and I've not found any variance that I think are worth testing in the wet lab. And then it's a business's decision to say, okay, great. Let's take the candidates that we have now into that next phase and call it a day.

Karl And so just, to be clear, like all of this iteration on the Cradle platform has happened computationally, and you end up with a protein design that has to be expressed or created.

Karl Cradle has also built, wet lab [00:25:00] capabilities and, and you're involved with doing partnerships. How do you balance both that scientific depth with commercial scalability?

Elise Scientific depth is incredibly important. Our scientific users are the absolute experts on the problems that they're working, on.

Elise So we have huge respect for their understanding and all the nuance of a system, whether that's a complex enzymatic assay or a specific mode of action that they're looking to engineer a protein four on our customer success team, which I, lead, we have customer or scientific advisors which help to translate.

Elise That project and all the nuance and all the complexity in it to the cradle platform. So to the world of machine learning models, that's where the scientific depth is important. And spotting the trends and helping to see, like on a meta level, how does this map to machine learning, but the context is equally important.

Elise So yes, we're working on this specific project, but how do you evaluate success? And now [00:26:00] when in an era where many organizations are. In their first year, of evaluating generative AI in their, large molecule development. there are many leaders thinking about how to evaluate success. Like Are we just looking for impact? Are we looking to reduce the number of projects that get stuck? Are we looking to speed up? Are we trying to show that different users in the different scientists in the organization can use these approaches effectively? are you comparing it to a traditional campaign like running those side by side?

Elise so that's where the commercial scalability is. like it helps to understand what does success look like before you go. Now we're gonna roll it out at scale because you have to be intentional when you do that, and you have to know what you're looking for and what good looks like.

Erum Yeah. And let's continue talking about this commercial side, right?

Erum I mean, you're selling to customers and like us at Messaging Lab, you work across biology and biotech. So the pharma side and the non-pharma side. [00:27:00] So I would love to know from your perspective, what differences do you see in how these industries approach innovation in biology and where do you see the bigger opportunity?

Elise There's huge opportunities in both the pharmaceutical and the non-pharmaceutical domain. my background is in the non-pharma and the industrial biotech space, where you see that it's really important to get extremely performant molecules. the cost of goods sold, is important. So there's a lot of competition on cost and getting squeezing the last remaining.

Elise Juice out of an enzyme like getting it just that improvement. Because if you think about giant bulk fermentation processes, like every, truck of sugar that you don't need to drive to such a facility is, you know, it's a win. Um, and you're competing on the cost of that final, product and what that translate to.

Elise In enzyme engineering, campaigns is that we, often see campaigns that are very long. They spend, 12 to 20 wet lab, and design [00:28:00] iterations, which is different from the, realm of, pharmaceutical. and drug development where there's a race to be first in class and antibody programs are on the clock always.

Elise You only have a couple of months. You only have a couple of iteration because otherwise you've already lost the race to, getting, your pharmaceutical out to the market first. these dynamics change a little bit how the projects are set up, the timelines of them, sometimes also the resources or the risk appetite.

Elise or the willingness to try different approaches in such programs, but there's opportunities across the board, and no matter whether you're using generative AI to optimize a peptide or an antibody or an enzyme, there is the power of these computational approaches that you can use. But you're always going to be feeding real world wet lab data that is specific to that project into the models to get the biggest strides to your final goal.

Karl That's amazing. and one [00:29:00] thing on this podcast we've talked about the Advance Biotechnology for Sustainability Coalition and Cradle's a founding member alongside L'Oreal Evonik Arsenal Bio Yards. And in the first a B eight forests report, they claim that advanced biotech could cut Global emissions by 5% and created a trillion dollars in value, I think by 2040. But it also says the industry hasn't lived up to its potential or its promise. What do you think the gap is between that, headline of a trillion dollars and where we actually are today, and what needs to change quickly to make that happen faster?

Elise Yeah. I'm so excited for the industrial biotech space, and I want it to live up to its full potential. What I notice as a deep structural problem is cost competitiveness, like bio-based alternatives are still often more expensive than their petrochemical equivalents. sometimes they are competing with subsidized ways of, [00:30:00] production as well, so production cost would need to fall.

Elise In line with, for example, the cost reductions that you've seen in other clean technologies like batteries or like solar, as these scaled up. So really want to see the costs come down. Um, you can think about how can we leverage AI to optimize the different parts of those systems. So optimize the proteins, optimize the strains selected, maybe building more effective.

Elise technology stacks that companies can use. So rather than everyone inventing every wheel themselves, how can we collaborate more across the industry to say, okay, we are a software vendor. use us alongside this DNA vendor, use us alongside this CRO who has this, race track for this specific type of protein to get the, high throughput data at short turnaround time to feed machine learning, and build those approaches together.

Lizette Okay. High throughput screening has come up a few times in this conversation, so let's go over it, shall we? High throughput screening is when scientists test huge numbers of biological [00:31:00] experiments at the same time, robots, automation and microplate allow researchers to test thousands of protein variants in parallel.

Lizette Instead of testing one idea at a time, you test entire libraries. Its biology running at industrial scale.

Elise So that's of our vision for. For the biotech of the future is that you can see what is out there in the ecosystem and use different innovations across the ecosystem rather than building every team yourself using your hard earned VC money right from the start. it's a, it's a future I would be excited about.

Elise The other thing that needs to happen to see this industry coming to its full potential is. Scale up. So more tanks built, more fermentation facilities, more pilot facilities. I believe the report mentions we need $500 billion of investment by 2040 to realize the tanks and, scale up facilities that we need to boost this industry.

Elise not all VC funds are set up for this type of capital intensive scaling, so. [00:32:00] What if we instead pulled capital into larger vehicles and found ways to, make the funding for the, these initiatives more viable to make sure that those innovations actually make that translation from the small scale lab into, the market.

Erum Oh, interestingly, isn't $500 billion the same amount that was required to create data centers for ai? It is like we just, that fight that that number is like, we just need $500 billion to do these big ambitious industry creating projects.

Elise Sounds about right. Yeah.

Erum it's like, oh, AI got $500 billion biology against my hundred.

Erum That too. We

Elise want that too.

Erum yeah, let's talk about some of the applications. I mean, you have experience in like, non biopharma or bio industrials. Can you walk us through, right.

Erum How did engineering a protein specifically change the way an application performed for a client? walk us through like kind of end to [00:33:00] end and, um, from the lab to. A product,

Elise projects that we've seen come in, have sometimes been ongoing for a couple of years and gotten stuck.

Elise So we've seen projects in which a customer came to us saying it'd be. Awfully nice to further improve this enzyme because it's going to be directly driving the cost of goods sold in this process. think about, for example, manufacturing an active pharmaceutical ingredient. So a small molecule using, enzymes as, biocatalysts to make this molecule.

Lizette Elise mentioned bio catalysis. That's when enzymes or biological catalysts are used to run chemical reactions instead of high heat, harsh solvents and energy intensive chemistry enzymes can do the same job under mild conditions. They're incredibly precise and efficient, which is why they show up everywhere from making medicines to powering laundry [00:34:00] detergents.

Elise but having seen in historic engineering campaigns that the gains across each. Next round of wet lab validation, we're getting smaller to the point of going, is it even worth pursuing this? so thinking that all the low hanging fruit had definitely been picked from the tree and, and potentially there were no, no apples left to pick.

Elise providing these data to machine learning models and allowing them to, identify whether there were additional. Variants based on, mutations that had been previously been explored that could be combined into better designs, but also allowing the models to roam throughout that entire protein molecule and spot new areas to mutate to further improve enzymatic outcomes.

Elise So performance of the molecule. and we have seen several instances where machine learning models were able to introduce mutations in. Parts of the protein that [00:35:00] scientists rationally had advised against or couldn't understand why that would even impact. And these are use cases where everyone sits back and goes, wow, that's, it's really cool that the model somehow came up with this.

Karl Yeah, it's amazing because I think that rationally, humans think that we would be able to. Also come up with those iterations and we probably can, it's just, we're just not as fast as the, ML models are.

Elise Yeah. And maybe to reframe that, we're not as fast as evolution either. and so like if Evolution took, however many iterations to come to this enzyme that works pretty well and is pretty stable and you, have a 96 well plate, like good luck.

Elise If you're just using rational design. and that's where it's, great to have models to efficiently explore that giant, protein fitness landscape and suggest, Hey, I'm finding a new peak in this valley and, check out this, uh, part of [00:36:00] the design

Lizette elise just talked about the protein fitness landscape. Think of it like a giant map of possible protein designs. Every point on the map is a different amino acid sequence, and the height of the landscape represents how well that protein performs. Scientists are trying to find the highest peaks, AKA, the proteins that perform the best. But the landscape is massive, so finding those peaks can take a lot of searching.

Erum Space. Quick question about that. In terms of, exploring this space, is it looking at proteins from.

Erum 2D structural level or three dimensions, because I know in the world of biopharma we work with, molecule design companies or small molecule design companies. And so they've been educating us on different types of data and how, there's smiles data, which is the 2D structures, like the different shapes.

Erum But then, so I'm curious to know, um, what data, guys are looking [00:37:00] at.

Elise As a platform, we use different types of models. There is a clear advantage in sequence models in terms of the amount of data that is already available in the public domain and how easy it is to create additional data and labeled data.

Elise So by that I mean, there is a couple of orders of magnitude fewer. Crystal structures in the protein data bank than there are sequenced, protein sequences in, sequence databases. And it's also much, much cheaper to obtain a sequence of, like a DNA sequence or protein sequence and. A label, so an assay value in the wet lab than it is to produce a new crystal structure, for a protein variant and its performance in the wet lab.

Elise So both types of models have a place in the protein design workflow. but if you're iterating and [00:38:00] making variations and, pushing across these multiple parameters, sequence models. do have the benefit of, receiving more data points at a lower turnaround time or more frequently because for us humans, it's less expensive and cumbersome to, run a sequencing run than to, try and crystallize a, protein variant.

Karl Yeah, that makes sense. looking ahead 10 years, what emerging technologies, whether in AI automation or synthetic biology, do you think are really going to most dramatically shape the way we engineer proteins?

Elise definitely AI guided protein design, and you see that that is being widely adopted.

Elise There's a lot of creative minds, both from within the bioengineering spaces as in, from traditional computer science domains. figuring out which new angles to take to explore, the usefulness of models and biological problems. I'm also excited for developments in data [00:39:00] collection because these are learning systems.

Elise It's machine learning. It needs something to learn from. it needs something to learn about. so assays that focus on. bringing, like bridging that translational gap from the very small wet lab to maybe those final, the application testing. so further miniaturizing assays, repeatably, so that you can run assays at that 96 Well plate throughput that is amenable to machine learning, rather than just, running your assays for one or two of your favorite, proteins.

Elise There's also innovation on very large, complex like ultra high throughput screening technologies like creative libraries and libraries that meet other libraries. And then coupling that to, sequencing technologies. And one thing that I'm also monitoring is what's happening on the DNA assembly.

Elise In the DNA assembly [00:40:00] space, 'cause DNA is, getting cheaper. you can order bespoke libraries already, but what kind of control do you have to say? I'm running a machine learning model. I'm generating a giant library of sequences. Those are carefully selected by the machine learning algorithm as highly valuable points of that enormous sequence space to observe.

Elise now can I turn that into a highly bespoke. Library, to then screen for the function that I'm looking at. So what type of DNA assembly methods and, synthesis methods are there to support? this new feature, which is machine learning, can propose very distinct points in the sequence landscape that it wants to see, rather than maybe generating randomness or, libraries based on just recombination or random recombination.

Erum Yes, well, a lot of exciting things. Hopefully coming online soon to make biology more accessible, faster to use, or easier to use, let's just say that, or easier to train. And biology is its [00:41:00] own beast. It's live it's life. So to be able to, I guess, of course, engineer and to making what we need to sustain our lives, so,

Elise Yeah. And I, what I love about it is that it's going to take all these different types of scientists. We have this incredible crossover now from wet lab scientists to hardcore computer scientists all of a sudden working on the same problems and exchanging ideas and asking questions, and I think that's such a fertile ground for new approaches and rethinking how we produce data, what kind of data we produce, and with these new technologies also, what questions do we dare to ask?

Elise and what, problems can we now set out to address because we know that we're armed with a more powerful toolbox, both on the, computational design side, as well as on the wet lab validation.

Erum Yes. Yes. Absolutely. Absolutely. and we love the idea of convergence. This is, one of the reasons why we have this podcast is to be able to bring all sorts of voices to talk about how we can leverage biology.

Erum But now we are at the part [00:42:00] of the podcast where we have our quick fire questions, and we're gonna ask you, are you ready for this? Before we get into it,

Elise I'm, I'm ready.

Erum Okay. Okay. So we're gonna ask you a series of questions and I'll go first and then Carl will, ask a question and then, just quick answers.

Erum Okay. So here we go. If CRISPR had a theme song, what genre would it be?

Elise Funk.

Erum Funk. Yes. I love that.

Elise Happy funk. What can we make work?

Erum Happy funk. I love it.

Karl What's more stubborn? A fungal strain that won't express or a buggy line of code.

Elise That one's easy. A fungal strain. Science comes with wonder and humility because in the end you're studying this complex system.

Elise Trying to alter how it works in places, which is different than the world of software engineering, where in the end you're creating your own universe and it's a ni nice and deterministic.

Erum Yep, yep. Absolutely. So if you could [00:43:00] engineer one everyday object to be protein powered, what would it be?

Elise There's octopus that live in shallow water and in their stomach they've got these bacteria and they grow during the day. And then. Come evening, there's enough of them. And then they use quorum sensing to all become, bioluminescent so that this octopus can hunt in shallow waters without having a shadow.

Elise So not scaring off the fish. So some sort of like bioluminescent lamp that is off during the day and then turns on in the evening. that'd be pretty fun. I'm also thinking of a protein powered coffee maker. I haven't thought through what that would look like.

Erum Oh, I love bioluminescence. So you really hit the nail on the head for me with that answer.

Karl Yeah, and I love the coffee, so that's gonna be a good thought experiment. what's the biggest myth about AI in biology?

Elise The biggest myth about AI and biology is that it's too hard to use for scientists not trained in computer [00:44:00] science or that it requires more data than. Is realistic to ever produce in your own lab. We are now at a point where generative methods are so powerful that they can learn from smaller amounts of wet lab data that you can produce with relatively accessible, modifications to a lab.

Elise If you think about. An open trons liquid handler or a minion nanopore sequencer, it's a lot more accessible to develop robust hydro throughput, wet lab workflows to feed machine learning. And these methods are becoming more usable, but you just need to know when to use them, and what types of problems they're suited for solving.

Erum Awesome. Okay, final question. Most underrated microbe in the lab.

Elise Oh, I'm gonna do a shout out to Calvarium, which I think has, is training a whole zoo of special microbes to work well in [00:45:00] the lab for all these different types of purposes. Personally, I mean, fi nitrogens so fast, so fast and we're not using it enough, I think.

Elise so it shows that we're, it's easy to get stuck in your ways as a scientific community, and it's, sometimes beneficial to rethink and maybe sketch out what the future might look like and actually build towards it and implement it.

Erum well said. Yeah. Verium, I mean, Henry Neely, hope you're hearing of you listening, because this is for you and we've been trying, we had Neely on the podcast and she's amazing.

Erum Now we're trying to get Henry on. Henry's been kind of slow to book his slot, so this is

Elise Okay. Book your slot, Henry.

Erum Oh my gosh. Yeah. So this has been an amazing episode. Thank you so much for bringing us into the world of protein engineering and design. There's so many tools that have converged to enable Cradle to be able to do this [00:46:00] and to shed some light on it and share your world with us. We really, really appreciate it.

Elise Yeah, it's been a pleasure and thank you for sharing the Wonders of Biology with a broader audience. to all you computer scientists out there, you're welcome to join us here in the wonders world of biology.

Karl Thank you so much, Elise. Really appreciate it.

Elise All right. Thank you both.

Karl Iran, what did you think of that interview?

Erum Oh, I loved it. You know, from being a [00:47:00] scientist, I had to put my scientist hat back on. It took me back to my lab work days. And even when, back in the day when I was in the lab, there was so many points throughout experimentation that I was like, oh my gosh, I wish we had some type of automation or computers can help.

Erum Design the experiments better. And this conversation really took me back there. But I loved it. I mean, I do love the fact that she gave us a point of view of what's going on, not only on the. Lab side, but on the computation side, and we talk about that a lot.

Erum We bridge the two worlds. because that is what's happening now and how do we leverage all of the new technology to create really fundamental differences to how we approach biology.

Karl Yeah. And what was great is, she gave us some insight into her experience. She had been at, a large food engineering company and then had been at Zymergen.

Karl These are names that people [00:48:00] in, the synthetic biology industry know well. And, that combination of her experience plus this need to be able to accelerate. The design and engineering of proteins and how close she was to that problem I think was really, super enlightening. And I think, like you just said, Iran, it's like we're beginning of, we're at the beginning of the shift where biology becomes a design problem, and that's very similar to the way software change what's possible and communications and commerce.

Karl Engineered proteins are gonna change what's possible in health, food, materials and manufacturing.

Erum Yeah. And then that's really fundamental to what we talk about here on the podcast.

Erum cells can theoretically produce almost anything, right? So fabrics, food, fuel, medicines, and better protein design tools could make that vision a practical reality rather than a distant dream. So there's a lot to be done. I'm excited to continue to learn about what Aise and the team of Cradle are doing to [00:49:00] support the development of the Bioeconomy, which products are really.

Erum better off because they had better protein design. And so getting those specific stories really helps ground how at least Cradle and these companies that, you know, we call it very upstream, right? It's like, okay, so you really looking at a small protein, it's on such a small scale and to go from that to medicines on the shelf or products that are on.

Erum medicines at the pharmacy, and products on the shelf. it's exciting to see how we can optimize and have better products for us that can help not only us as humans, other animals, the planet. We're here for it.

Karl Yeah, and what this reminds me of is a few episodes ago we had Joshua Lacer of Synonym, which is now known as Robing, and Joshua had talked about the Great Reformulation, which is about the reformulating of all consumer [00:50:00] products.

Karl And what proteins make possible is not just swapping. One ingredient for another, but really designing ingredients that, for example, a protein that has an activity like cleaning your counter can also have like an antibacterial or a preservative effect so that you could put less ingredients into a products and they can be more effective.

Karl And that's really what protein engineering like the cradle team is doing. And it is what the Cradle team enables. So I'm really excited for that.

Erum Great episode. You know, we love having scientists come on sharing their world. We're excited for some upcoming episodes and if you wanna meet Cradle and other companies that are building the Bioeconomy, you have to come to Sin by Beta this year. It always happens in May. And for anyone that wants to come. We do have a discount code, use code, grow everything, to get a discounted attendance ticket.

Erum We would love to see you there.

Karl Thanks for listening. Please follow us on YouTube or any other channel [00:51:00] that we stream on, like, and subscribe and We publish notes on the episode and previous episodes on.

Karl We love hearing from our listeners. We have a 800 number. You could leave us messages. We take those messages very seriously. That's the pod.