AI Can Now Write DNA. What Does This Promise, And What Could Go Wrong? With Eric Nguyen, co-founder and CEO of Radical Numerics

0:00Eric Nguyen:What's really exciting or was really exciting last year that folks really took us to the next level and it was part of our initial goals which was creating these gene editing tools which are very complex and very useful and interesting. The most ultimate machinery is in tests is can you create a whole organism from scratch and indeed scientists actually did use Evo to create the first AI genome so a complete set of DNA for what's known as a bacteriophage. They spent two years on very painstaking experiments. Our teammate spent 30 minutes using our model to be able to rank the DNA from these subjects and identify this and rank the same genes that were most causal to Alzheimer's.

0:41Eric Nguyen:It's all based on the spelling of the DNA. So they'll take a sample, they'll sequence it, and basically take the letters and match it to a database. Now, a deepfake virus, there's a line of work that showcases that it's possible to use AI to essentially change the spelling. Welcome, humans, to the Neuron AI Explained. I'm Corey Knowles here with Grant Harvey. How are you today, man? Doing good. Doing good. And I'm especially well today because we're talking about a really interesting topic. So most generative AI writes text, image, or maybe code. Well, today's guest, Eric Nguyen, helped create Evo, which is a model trained to understand and generate DNA.

1:23His new company, Radical Numerics, is building even more capable systems for genomic research, disease detection, and biodefense. I want to learn more about that. Today, we're going to unpack what these models can actually create, how close they are to producing useful new medicines, and whether our ability to detect AI-engineered biological threats can keep pace. Before we get started, please take just a quick second to like today's video and subscribe to the channel so you never miss an interview or one of our live streams. Real quick, I want to take a moment to thank the sponsor of today's video, BeyondTrust's AI agent security.

1:56You can learn about them later on in the episode and in the link in the description below. And on that note, Eric, welcome to the Neuron. It's great to have you.

2:05Eric Nguyen:Great to meet you guys. Thanks for having me on the show. And very excited to talk about what we're building here at Radical Numerics. Excellent. Well, I guess let's start really simple. When you say that an AI can read and write DNA, what is it actually doing and how is that similar and different from how ChatGPT reads and writes language? Yeah, great question. So actually, in many ways, it's very similar to how ChatJPT works. Conceptually, what we're doing is treating DNA also like a language. And in many ways, it is like a language. It's got sort of grammar, structure, in terms of vocabulary, but it happens to have a very different set of letters, right?

2:53Eric Nguyen:It's got these letters A, C, T, and G. But those are just really stand-ins for what they represent, chemical compounds. But they do have a certain way that where you have them basically ordered in a specific sequence, it changes its function. It changes its properties. And so it makes those changes basically following a set of rules. And for humans, it's just really been hard to understand what those rules are because, well, it isn't natural language. And so the hope is that using AI, we can essentially train it to understand these underlying physical rules. And ultimately, if we can understand, then we can control.

3:33Eric Nguyen:And that's going to help us understand how to treat disease and ultimately improve human health. For people who are a little more technical, is this a language-based model or is it an entirely new type of model? Yeah. So it is a language model. And so large language model, just like other types of AI. Now, how you get it to work on DNA, that's sort of the trick of the trade, because it is a different type of data distribution. That being said, what we, our team is known for doing, we, the founders of Radical Numerics, there's four of us, we met during our PhDs at Stanford doing AI research. And we were working on essentially novel architectures, new models to handle sequences, not just language, but novel types of sequences.

4:25Eric Nguyen:And what our models were particularly good at, we called this model at the time hyena. It's a new type of architecture. It was very good at long context, which now is a very common vernacular that people are familiar with in language models. How long are we talking? How many tokens here? Yeah. So back in 2023, the longest context models at the time for DNA and language were on the order of a few thousand. Like in DNA, it was like 2 ,000 tokens. Wow. We had created the first language model that could handle a million contexts. Wow. Yeah. And so now our newest models are now at two million. But back then we were pushing the envelope already to a million.

5:10Eric Nguyen:And that was about 500 times longer than previous language models on DNA. And so what we started to get people to feel that was possible was this idea that we could train AI to understand DNA, which is basically the longest sequence of them all. It's the longest context you'll ever find, generally. What's like, just for context, what's the average DNA strand? How long is that? In an order of magnitude, let's say. Yeah, so the best example is us, right? So we are made of DNA ourselves. All our cells have DNA, the instructions to encode an organism. That is 3 billion base pairs. They're called base pairs.

5:55Eric Nguyen:Base pairs long. And so the way we tokenize and feed our language models DNA is one token equals one letter. And so that means you have a 3 billion token context that you ultimately want to be able to read. Yeah. Wow. So I'm curious, what does the model understand about, like, biology as opposed to just thinking of, like, predicting the next likely sequence? What does it understand about the world around it? Yeah. As you said, so it first basically learns like language models. It's like this, you know, think about next token prediction or in some cases mass language modeling. But through this pre-trained self-supervised task, what it's able to basically learn is this implicit understanding of the rules and grammar of organisms in life.

6:53Eric Nguyen:And so we've trained it on DNA across the tree of life, like pretty much every type of organism on Earth. And there's a lot of shared information and information about what an organism is, what it needs to survive, what is conserved, meaning what is basically most useful for survival and what's passed down. and specifically for us, we can take advantage of those implicit things it's learned and essentially you can think of it like prompt engineering the model to be able to generate new stuff, like understanding the rules and grammar of DNA. Can you use that to then essentially autocomplete new functionality or new organisms with the hope that you can essentially create medicines or treatments for people that did not exist before.

7:48Eric Nguyen:Wow. That is so cool. That is. What's a good example of one type of organism that you're targeting or one type of, yeah, like what are you trying to do with that first, I guess is a good way to put it. Yeah, absolutely. So starting with the maybe chronology of some of the interesting things that we were able to create with Evo, the first example was being able to generate the first CRISPR-Cas system using AI. And so CRISPR-Cas is this gene editing tool that exists in nature. but also there are many more in nature that we haven't discovered yet or we might want to come up with synthetic ones that have different functionality and being able to they're like molecular sources they can edit dna and what we're able to do is essentially train evo on a bunch of existing crispr cast systems and then essentially ask the model all right given what you've seen before, can you generate something more efficient or efficacious?

8:54Eric Nguyen:And so essentially by prompt engineering the model, it's in many ways like hallucinating or like coming up with new variants. And then the scientists that we're working with actually took those designs and synthesized them and physically created them in the real world and tested their functionality. And indeed, they were able to cut DNA. They're able to work and function like editing DNA, just like other types of CRISPR-Cas systems. So that was the first. And I think what's really exciting or was really exciting last year that folks really took us to the next level. And it was part of our initial goals, which was not just creating these gene editing tools, which are very complex and very useful and interesting, but sort of like the most ultimate machinery is and test is can you create a whole organism from scratch?

9:41Eric Nguyen:And indeed, scientists actually did use Evo to create the first AI genome. So a complete set of DNA for what's known as a bacteriophage. It's basically, in other words, a virus, which when folks hear about that, a couple things come to mind. You would be like, no, we should not be doing this. This is what Daria is always talking about, like trying to stop. Yeah. So naturally people have this reaction, right? Yeah. Yeah. And so, yeah, peeling back a couple of layers, bacteriophages are harmless to humans. They cannot attack humans and they actually have beneficial use, meaning because they attack bacteria, they can essentially be used like antibiotics.

10:25Eric Nguyen:And so they do have beneficial use cases. So in this particular case, what's exciting is that there's a potential to create these precision custom antimicrobials. microbials. And I think that was exciting for the scientists to showcase that, hey, we can create something that did not exist in nature. It has human benefit. But also, right, there is this inherent threat where, yeah, if one can make such an organism, have such control over the manipulation of life itself, right, then the idea of creating something harmful becomes very top of mind for folks, right? And I agree. In many senses, being able to lower the barrier to make biology controllable, programmable, inherently has this inseparable ability to make things that are dangerous as well as beneficial use.

11:22Eric Nguyen:And so that's why as a company, we actually have a dual mandate for us to push on the capabilities of biological AI design, for the beneficial use, but also make sure that it's used on the defense side, on the biodefense side. And this was a key turning point for us, this being able to generate a whole genome, in this case a virus, that we saw this ability and also opportunity to apply these models and get ahead of it and make sure that these are safe in biodefense applications as well. Okay, quick security reality check. AI agents inside enterprise environments are growing 40 % year over year, and 7 % of organizations already had an AI agent-related security incident in the past year.

12:01The danger here isn't that agents can invent some new permissions. It's that they can weaponize the permissions that are already sitting there. So an agent may launch with the same cloud keys, SHH keys, and login tokens as the person running it, often with no scoping, no exploration, and nobody watching. That's why BeyondTrust built AI agent security. It discovers every agent running in your environment, including shadow AI that nobody approved, and traces each action back to the human or agent behind it. Then it steps in before damage happens by blocking destructive commands, limiting credential access, and requiring human approval for sensitive moves.

12:36One policy works across Cloud Code, GitHub Copilot, Cursor, and more. So be among the first to secure your AI coworkers before they act. Visit beyondtrust.com and learn more about Beyond Trust's AI agent security. I'm curious. I assume the long-term approach. to being able to make these kinds of changes and edits to someone's DNA, uh, is, is to give better resistance to maybe cancers, maybe other diseases, maybe aging. Uh, but what I'm wondering is, is this an approach where this is editing a piece out, replacing it? Is this adding a new piece in that strengthens resistance for it? Is this, uh, How does something like that even work?

13:25I realize that's maybe a little grammar school, but I want to make sure I understand.

13:30Eric Nguyen:It's a great question. I think this is such a foreign concept for folks that may not be in the sciences, and so totally reasonable. I think what's first kind of setting the stage to how these models work, they focus on the design side, right? So these models are generating sequences and helping to control or program functionality into DNA. And so how you use that, that knowledge, that design capability can take many forms, right? You can make DNA kind of like standalone that's like separate from your body. And you can think of that more like a drug or a separate biological machinery on its own.

14:11Eric Nguyen:And then also you can make something all the way to a whole organism like that. There's a huge range. And so that's one category which is kind of like separate from our bodies, right? So like not on our raw bodies. And then there's a whole category of, well, if you want to manipulate the DNA of something like an animal or a person, that's a whole other category of like gene editing, right? So then that's a category where we as humans now have the ability to edit our DNA in vivo or in a cell or body. I think where AI interfaces with that ability, right? Because AI doesn't have the physical components.

14:52Eric Nguyen:But what it can do is tell you if you make these edits, what are all the cascading side effects that can happen or that you can anticipate? Right? Because that's the biggest fear. One of the biggest fears for gene editing is that if you make an edit where your intention is to cure a disease of some sort, what are the unintended consequences that cascade and that you can't account for. And so being able to know what edits to make and how it plays out is what AI can help with. That's awesome. Yeah, that is so cool. And is that something that we can, I guess it's something we can't really fully know at the DNA level, but we can get a lot closer to knowing it, right?

15:35Because I assume at some point you're going to have to simulate the entire body and how it works together as a system before you can really predict all of those consequences. And each unique body is different, right? So we probably have to do it for everybody at some level. This might be a naive way of thinking about it. But how are you thinking about that problem of like, can you handle it all at the DNA level? Do you have to model things at the systems level? How do you think about it?

16:03Eric Nguyen:Yeah, such a great question. and actually a really big motivator for what we're trying to create at Radical Numerics. So, and it's also a little bit of a philosophical question, this idea of like, is DNA all you need to know to understand everything about your body? Actually, some people say yes, some people say no, right? So this is actually an open question in some ways. For us, I think DNA is the foundation, is the starting point. But DNA is not in a vacuum. So it's interacting with the world. It's interacting in your body. And how the DNA is used in your body, in different parts of your body, in different time periods of your body, this is sort of measured in different – what I call different modalities or different ways of manifesting.

16:55Eric Nguyen:Manifesting like the right kinds of proteins that are made from certain genes from your DNA. Gene expression is changed by a lot of different other physical factors around your DNA. And these are captured very differently than, you know, somewhat similar to DNA, but in different contexts. And so what I see as helpful in understanding is how do we provide that extra biological context with DNA so that we can understand how the complexity in the system of our bodies, how it behaves when a disease happens or when a drug is taken. And being able to model all of these other functional properties has been what's missing for some of these DNA-only models.

17:44Eric Nguyen:And so what we've done as a company is essentially figured out how to tokenize or get into a form that's useful for large language models to be able to read those modalities. And so specifically those modalities are like proteins, protein structure even, like RNA, epigenomics, which is like the really environmental kind of measurable things around your cells. Metabolomics, all of these things are basically, in my mind, somewhat like different languages that have been sort of learned individually on their own throughout science and biological history. But now for the first time, we have this ability with potential with AI to not model one specialty or one siloed modality at a time, but to integrate it all into a single system that can actually measure the complexity of something like our bodies.

18:37And that's what we're building toward. And I think approaching it from a long context perspective is the key because to be able to do that, you need to be able to hold so much context in your head. me anthropomorphizing the data centers here or whatever, like the models. You need a data center in your head, Joe. Yeah, you need to hold so much information together at one point. I feel like$3 billion is just the starting point for how much context you need to do that.

19:06Eric Nguyen:Yeah, yeah, absolutely. Yeah, so we like that analogy of like, oh, it's different contexts, especially for biology. Like people think of context usually in these language models in a similar way. Like the more context you feed to a language model or chatbot, you know, the more specific it's able to answer your questions, the more it can adapt, right? Like, oh, for example, if you ask a chatbot to like write you a story, let's say, if you ask it to just broadly write you a story, it'll write something, but it would just kind of go for, you know, a bunch of directions. But if you ask, if you give it more context, like, oh, I want a mystery story set in London in the 1950s, you're providing more context and it's able to generate things that are basically following that regime, that pattern.

19:49Eric Nguyen:And in similar ways, what we're doing on the language models for biology is also feeding additional context, both for DNA, like sequence lengthwise, but also in these other modalities. So like give it context for a specific cell type. Give it context for a specific disease state. And that also shows up in different similar sequence information from biology. It's like it's not language. It's like this physical world property kind of stuff. In many ways, it can be kind of thought as like time series kind of cell data for models. But anyways, essentially for a lab, an AI research lab at Radical, we've figured out how to tokenize and discretize the data so that it's fed in a native way to these language models, that it can actually learn all of these contexts so that it's like in a certain disease state.

20:38Eric Nguyen:It understands what the methylation patterns are for someone who has Alzheimer's. Alzheimer's. And another person who has cancer, this is what the shape of their DNA through chromatin accessibility looks like. And it's shown up in these modalities and different phenotypes, molecular phenotypes through sequence information. And so being able to feed those different contexts, you now have a big context adapting machine to these different states of biology. That's what I was wondering is I think the idea of how do you pre-train a model on the volume of DNA it would have to take, I presume, in order for it to be able to do what you're doing.

21:24I mean, like, that has to be just a monster undertaking. And simultaneously, I can't help but wonder, is, like, any of the Human Genome Project stuff, is that a thing that enables this in some way that was done over the years?

21:38Eric Nguyen:Absolutely. So those are a great question. So the Human Genome Project was really one of the first steps for this endeavor, right? The Human Genome, what it was trying to do was basically trying to come up with the actual sequence of letters that makes up the genome. And we thought, you know, as humans that once we could do that, this is going to unlock. We've unlocked the world. We've unlocked the world. We're going to solve biology. We're going to understand everything that there is about. We've won biology. Yeah. We've solved. It's solved. Yeah. And so I think what was, you know, it was extremely successful in many ways.

22:16Eric Nguyen:It has opened up a lot of directions and made, you know, a lot of treatments possible and understanding of biology far possible. But it did have its limits, right? Yes. It's able to show what letters are. You can physically, like, read what letters there are in our genome and the exact orders of the letters and combinations. But being able to interpret what those letters mean was really the next step and still is a giant step for what we need and have been doing to really understand human biology. And so that's been the challenge left open. Now, being able to train and hopefully use AI to live up to that challenge, as you said, it can be in many ways a very big undertaking.

23:06Eric Nguyen:And one of the reasons is that the scale is a different scale in terms of data and context length. But in many ways, it's sort of similar language in the sense that just, for example, DNA, looking at DNA publicly available online. There's more DNA online than all the text on the internet. So it's like – it's internet-scale data. Whose DNA is this? So it's not necessarily human data. There's human data, but mostly it's other species. So it's the tree of life. You know, organism, microorganism, mostly microorganisms actually. Mammals, plants, fungi, archaea. So it's the diversity of life you can capture online, essentially.

23:56Eric Nguyen:It's not annotated. So people also in many ways don't know what the grammar is, right? And so one of the hopes is that being able to use a language model, you can distill a lot of that underlying grammar into the models. And then the challenge is then how do you figure out how to do something useful with it? And that's a big challenge for us and an opportunity for us as an AI lab to be able to harness what these models can learn and ingest, but then be able to tease out what it's learned and do something incredibly useful and beneficial with it. That is the name of the game. Yeah, because if they could annotate it, even just the existing data, if you can use AI to annotate the existing DNA that's known.

24:45yeah it's unlocking a whole other level if it could just go back and add definitions to everything based on everything it's learned this is what this is, this is what that is that would be pretty cool give me an index then you can see all the patterns across everything and then maybe you do have a key to unlock the tree of life yeah

25:04Eric Nguyen:this is an active area people do indeed do this there are companies a non-profit in particular comes to mind that works on annotating DNA from nature using AI models, ingest and be able to feed in annotations of known things and leverage that to annotate the world, annotate the biological world. Very, very ambitious endeavor. What's your philosophy as far as the future of how far this goes at the language model level? Are you imagining like one big God model to kind of like, like, you know, kind of like the anthropic model where you train just like a massive, massive size model and then slowly you distill that down?

25:49Or do you think that there would be a future where there's lots of smaller models, like you make like a model specific to the organism that you're trying to create? Maybe we make a human model. Maybe we make a butterfly model, whatever. And then even down to the point where could there ever be a model of me so that my doctor could essentially have a model that helps knows everything inside my body? Corey knows that too. Maybe it's a model that if you go the God model route, maybe it's a fine tune on top of the existing main model. But what's your philosophy on how that's going to play out? Yeah.

26:28I think what's incredibly exciting for biological sequences is the scaling laws that we're seeing is exhibiting similar emergent

26:44Eric Nguyen:capabilities as language. And I don't think people know it for language necessarily either. I think people have been scaling obviously constantly right now and still are. And people still wonder, is that going to be tapped out at some point? So, like, ultimately, I don't think people do know what the limitations are. It's just that current trajectory. It seems to be. Keep going, right? I think there's a similar sense in bio that – well, actually, I'd say in many ways people think that things don't scale in bio. But I think what we're going to showcase as a lab very soon, hopefully, is that there is a lot more scale that people have left on the table.