🔬 Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

The design side's gonna get more capable. The defensive side needs to try its best to get ahead. So I think inherently there is this arms race style dynamic that the defensive side has been far, far lagging. And so what we wanna do is bring the defensive side to par, essentially. We felt it was important as a lab that a team that was both building the design capabilities is actually also best suited for building the defense capabilities, 'cause they're basically the same models. Yeah.

A model that is good at generating, turns out is also very good at discriminating or predicting if a sequence is pathogenic or not. For, for us, we as a company thought it was very important to have a, a dual mandate. It's this idea of essentially being cognizant and, and feeling responsible for the capabilities that we're enabling on the design side.

So if we're going to create models that can design function into sequences, we believe and see a gap in companies being able to safeguard that technology. Welcome to Lane Space. I'm Brandon. I build RNA Therapeutics at Atomic AI. I'm joined by my co-host, RJ Haneke, uh, CTO and f-- co-founder of Mirror Omics. Today, it's a pleasure to have with us Eric Gwyn at, um, CEO and co-founder of Radical Numerics.

Eric, uh, started-- got his PhD in Chris Ray's group. Uh, he spent a lot of time thinking about how to do long context genomic models before, uh, long context or genomic models were cool. He was the, you know, first author, and I think, you know, basically visionary behind the Evo generative model, one of the first generative genomics platforms, uh, developed Evo Two, which, you know, naturally led into, um, you know, Radical Numerics.

Uh, yeah. Thank you for being here. Did I miss anything? That sounds great. Yeah. Cool. [laughs] Welcome. Thank you. So, Eric, let's talk about, uh, Omni and the blog post that you guys did about the benchmarking. But I wanna hear first, okay, what is a genetic language model? Why do I care? What does it do? And then let's talk about the top line results from the, the blog post.

So, uh, a, a genome language model or GLM is a large language model trained on DNA sequences, so very much like natural language and chatbots you see, but not trained on words or natural language, but on the raw fabric of life, which is the sequence of letters that make up DNA. And we ourselves, our, our company, our team is known for creating the first, uh, generative, uh, genomics models, which are models trained on DNA, not just to read, but also write, meaning able to generate new sequences of DNA.

And we felt this was, uh, an area that was overlooked, and that if AI could read and write DNA, it could change a lot. Under-- you know, scientific discovery, uh, and understanding of human health and how to treat it. Um, and so we felt there was a big opportunity to train AI on, on the genome. What kinda things can you potentially do with, with a model like this? Great to start.

For, for, for us, when we first started working on DNA models, we worked on this model called Hyena DNA, which is a large language model, but it used a convolution instead of attention, so a little more technical details. Um, DNA has this property that, well, it's very long, right? Um, at the time, these large language models had limited constraints on, uh, context, right? Being able to f-fit long sequences.

And so we were looking for a more efficient algorithm to be able to handle something like DNA. And so we came up with this, what we call the Hyena operator, uses convolutions. Long story short, it let us process longer sequences, in this case, up to a million, and at the time was the largest context for a language model.

And what you-- what we did with it was essentially used it to read DNA, predict function. So given a sequence of DNA, string of characters, um, we predict its regulatory function, its effect on a genome. And, you know, this is interesting to scientists because a lot of the DNA in our bodies, um, you know, perhaps people are less aware, but actually we don't know a lot that much about our genome.

It's-- We know it, obviously, it encodes the information for making us, us and help, you know, all the different complexities and potential diseases. Uh, at the same time, the grammar, sort of grammar rules about how the combination of those letters are sort of formed, what they encode, and how they encode function and traits is not fully understood.

And so the hope was using these DNA models, language models, to be able to map some function from the raw DNA sequence. And so our first generation models was able to show that, yes, we can, um, train AI to be able to read and understand to some degree, um, DNA sequences, and especially, um, what we call the longer range interactions, meaning over sequences.

You know, if you use chatbots, for example, like if you've heard of the con-- the, the phrase, uh, context rot, you know, the longer the input you put into [chuckles] a language model, um, it starts to deteriorate. And so, um, being able to pick up long range information and sort of patterns, motifs, grammar over long sequences was what we were trying to accomplish.

Um, so we showcased, uh, in that first generation of Hyena DNA that that was possible over a million context. And then really what started the field now known as generative genomics, was the model called Evo. And Evo, we sh- tried to showcase there was this idea of not just reading DNA, but it being able to generate it.

And so, um, we wanted to re-- accelerate essentially how biologists and scientists had learned from biology and, and particular genomics. Um, and we felt like this whole field of, of generative AI being applied to language, great, accelerated the, you know, obviously our understanding and, and ability to manipulate the natural language.

Um, but here's this other language- DNA and the genome that we don't understand, and it's barely being applied with AI in, in our opinion, at the time, a few years ago. Models that we saw were really small, short context, so they could only pick up small patterns and limited context, and none of them generated DNA. So they all just would read.

And we felt that the idea of generation was so powerful and transformative in natural language, what if we could bring that to biology and, and DNA particular? What can you accomplish that you can't do in a lab, right? So what does that unlock for you if you could do that very well? I think one of the, um, first things that we showcased that got folks sort of intrigued by the potential of this was, um, a CRISPR-Cas system.

So it's an enzyme that's able to cut DNA itself, and I think what was particularly enabled by the Evo models was the ability to generate over not just one modality or one type of sequence, um, but spanning multiple modalities and spanning multiple scales. So, um, CRISPR-Cas, uh, it's, it's a molecule made up of both RNA and, uh, D-- and proteins. And so at the time, you hadn't really seen models that can generate multiple modalities.

You can s-- they had protein language models that can generate proteins. Sometimes you had RNA models that could pr-um, generate RNA, but you didn't have a single system to sort of co-design. We showcased that, um, a single DNA model, sort of the, the foundation of, of both of those, right? From, from DNA, you can get RNA and proteins, that we can design a single system to generate and also function in the real world.

So we asked Evo, we showcased it a bunch of natural, uh, CRISPR-Cas systems, and essentially asked it, "Can you, can you make a new one?" And we're able to sample from that, that model that we trained, and indeed, we showcased that, um, Evo was able to discover a new CRISPR-Cas system, and, you know, folks were, were intrigued by it and cover of Science Magazine, and later get to give a TED Talk about the work, which is interesting because obviously the audience is very general, and so trying to make, you know, what is CRISPR-Cas system?

Yeah. What is DNA, and how do you generate, why, why would you generate it? All sorts of fun topics. Um, and I think the more even intriguing, exciting thing that scientists eventually, um, just last year showcased what you can do with a generative DNA model was to generate the first genome from scratch using AI.

So, um, this is something not possible by humans, right? Humans usually, you can think of it like copy and paste parts of other genomes or other DNA- Mm-hmm... put it into something else. Um, but they would just take out small motifs that, you know, they know the function, um, they understand the rules. But to build something from scratch- Yeah...

from the ground up, um, had not been done before at the genome level. And so, um, Evo, turns out, was able to generate a functional genome, and in this case, it was, what's known as a bacteriophage, also known as a virus. Um, and this, this was a key turning point for, I think, the scientific community and for us as a company at Radical Numerics, because we felt this was such a, an indicative manifestation of the potential, right, to create a whole organism not existing in nature, but also the potential harm that that means as well, right? If you can control-- If you can manipulate the fabric of life itself, control its function, what kind of implications does that mean?

What, what are you enabling into the world? And so we actually got a lot of feedback, a lot of comments, a lot of outreach from folks, both excited and concerned about this capability, this kind of capability, and, and just the trajectory, right? It-- This is the early stages, the first thing one can sort of project and imagine what this could lead to.

And so we felt, as a company, important to-- it was important to not only push on the biological design capabilities of these models, but also the ability to use them as defensive tools, um, for the potential of, uh, misuse and, and biological risk that emerges. Um, and I think indeed, a lot of companies, a lot of frontier labs are also being concerned about this emerging risk of AI models being capable of designing biological sequences.

And, um, you know, at the same time, it's being mostly attacked from, like, a natural language standpoint, like safeguards and things like Claude. You know, if you talk about viruses, it'll just, like, shut you down. Um, which is great. I think it's, it's-- to some degree, you, you need, you know, safeguards at the natural language level.

But I think, uh, what you also need, clearly, is safeguards, um, at the biological sequence level too. So you need models that not just can understand language and, you know, the trajectory of, of your chat, but to understand the substrate itself is the next step and ultimate limit, right? If you can have models that can understand, you know, if the sequence is pathogenic or a virus, that's the level of, of defensive capabilities that you want.

And then, you know, being able to push that out into surveillance systems, um, national security of, you know, being able to monitor, you know, emerging sequences in the environment. This is what that kind of capability makes possible, and then we think bringing this frontier technology to, to that community as well is also important, just as important, um, as using this for human health, which is what we primarily focus on.

So- There was this evolution that you mentioned. There's the hyen-hyena DNA, and then there's Evo, Evo two, and now, uh, Omni. Can you just talk about, a little bit about Evo and Evo two struggled to beat sort of specialized models across many tasks, whereas in the blog post, you talk about how across a wide range of tasks that Omni is actually able to outperform them now.

So can we just talk a little bit about that? So yeah, Evo was, um, intriguing to folks in many ways, uh, showcased the, the potential for applying to multiple s- um, types of modalities. But it still, it still in many ways underperformed sort of the specialist DNA models, especially on human genetics or genomics.

And so although Evo was competitive, it, it still wasn't state-of-the-art or, or kind of pushing the needle. And so, you know, s-some parts of the community thought, like, "Why use a, a giant LLM when I can use these, these smaller, more specialized models?" And so what, what we wanted to do with Omni was a-amongst many things, but one of the first things was to showcase this idea of mid-training and post-training or, or broadly alignment.

So the way we think about language models in bio-- in bio and in genomics in particular, is that mostly we've only seen base models trained, so they're pre-trained, but they're basically unaligned. So in, you know, in the analogous space for natural language, it's like you're doing all the pre-training, but to make it actually useful in the real world and answer questions that users actually want and that is in the form factor they, they find actually, uh, informative, there's a bunch of alignment and post-training and mid-training done to get the models to, you know, be pr-production-ready and actually, actually useful.

And so we, we felt Evo was just showcasing the potential of that pre-training, but Omni is a step of actually making it useful for folks like scientists, right? And so, uh, we spend a lot of time on alignment and, and mid and post-training, which is essentially showcasing tasks in the form that people generally would want to understand about genomics.

You know, given a wild type and a, and a mutation sequence, you know, help, help me find the causal variant, right? These types of questions and form factors for how you might want to analyze genomics, um, doesn't just emerge necessarily easily on its own from pre-training. Pre-training is, you know, this next token predictions task or, or infilling.

And, um, basically, I think of it as like the raw, uh, pattern-making ability that you're, you're teaching it is in the pre-training, but then taking those learned embeddings or features, um, and pointing at specific tasks or, you know, a bunch of tasks really, um, and aligning it, meaning have it show you the output in a way that is meaningful to you, takes a little bit of teasing and manipulating that so far have-- like the front return labs are the ones that drive that research, um, in the natural language community.

And so we wanted to bring a lot of that research a-and more, um, to genomics. And so I think Omni is a, is really just a preview to showcase that potential, right? And I think once you do that even just a little bit, you know, we were surprised that it did start being, um, state-of-the-art and, and pushing the boundaries, not just being effective of multiple tasks, um, just broadly like Evo was, but actually pushing the, the frontier of each of those areas, variant effects prediction, um, causal mutations for disease.

It could ac-- It c-- It could start actually being useful for human genomics, and so we're, we're really excited to share that w-with folks. And it was just a preview in that sense because we're, we're still actively training and incorporating additional techniques into the model, like additional modalities.

But I think we were basically too excited, and we wanted to get this in the hands of folks faster. And, um, you know, some of the feedback we got in, in early interest, um, lots of hospital systems, um, nonprofits that have tons of genetic data and, for example, they know there's some kind of condition or symptom for the patient, but they can't figure out which parts of the DNA are causing it.

So they've got these VUS or variants of unknown significance that, um, we're extremely excited to apply these models to a-and actually help diagnose a lot of these patients, um, is but one example of a, a real use application. I'd like to talk more about the applications in a bit, but I am curious just from a technical standpoint, like what does it look like or what does it mean to align a genomics model?

I can imagine with large language model, there's a sort of natural chain of thought. You know, RL, there's like a clear paradigm here. I'm curious, what does it look like for a genomic model, which is not maybe better described as something like fine-tuning- Yeah... in a different language? Yeah. Yeah.

I mean, I, I think in, in many ways one could describe it as fine-tuning, but then introducing, I'd say sort of the key components are the right structure of the inputs. So feeding them in a certain sequence so that the model is aware that a certain task is being asked of it. Um, so there's a mix of special tokens to basically you can think of as like if you're gonna do disease prediction for, you know, for disease A, y-expect this special token, right? Um, just kinda like a way to prompt it.

If you expect it to do, um, design, have another special token, and then showcase the examples kinda like in a chain of thought manner, meaning showcase a sequence of desired outputs and the trajectory of it. Um, this is a little vague sort of intentionally 'cause it's part of our [chuckles] part of our, our, our secret sauce that we're still developing, and I think over time we wanna showcase more and more of it.

But, um, in many ways it does mimic a lot of the natural language community. It's-- a lot of it is fine-tuning, um, but really it's also carving out specific data sets that you want it to focus on and then structuring the questions or tasks in specific ways as opposed to pre-training it. It really-- Pre-training is really just feeding it in and, um, everything in really, and just doing next token prediction or mask infilling if you're doing mask language modeling, modeling, and it has no sense of like this Q&A type structure where you have a question, you know, a prompt and then an output.

And you can think of mid and post-training as starting to showcase, given this type of input, I expect this type of output, whether it's score, prediction score, or a design, is largely the mid and post-training. The post-training also includes things like reinforcement learning too, but, uh, I think the bigger steps are, you know, introducing structure of like question and answering.

Do the models have multiple heads that you-- are task-specific, or are you training one set of heads or whatever that can answer multiple questions at the same time? Just change the input tokens or whatever. Yeah. Broadly, it's a little-- you can-- I think we're flexible on this. The idea, um, yeah, sometimes you need-- y-you can use different heads, but the idea for us is to unify more so.

Um, and so I think early experiments, we did have different heads, but in some cases, the different heads do better. In some cases, the unifi-- the, you know, a, a single head, a single model, um, does better. And so I think we're flexible on that. But the, uh, I think broadly the direction that we are moving toward, um, is a single. And the reason for a special motivation for that, um, is that we're trying to unlock a lot of modality, try-- a lot of generalization.

And I think the more unifying, um, we're able to make these models, I think that's when you see more emerging capabilities, um, happen. And in, in large part, that's what motivated the D-the DNA work. We felt like a lot of models were specialized into other modalities, a little more downstream from DNA, so like RNA or proteins or small molecules.

We largely felt DNA is the foundation and that from DNA, you can learn a good deal of other modalities, um, ult-- potentially all of them. And I think of, of other modalities as sort of like additional context, um, that you're showing the model. That's how I kind of view it, uh, philosophically in my, in my head.

But yeah, the, the idea that single models unifying across modalities, scales i-is what the lab builds and builds toward. Can you walk through a few of the tasks that, that you talk about in the blog and, and just explain and remembering to, to, you know, narrate for the listener-only audience, um, but talk about some of these top-line results, and we can maybe dig into them a little bit. Sure.

One of the areas that we thought was really interesting for showcasing in, in this particular release for Omni was on understanding variants and their effects, which are essentially i-in DNA, a change in a position, you know, changing the letter of one of the other f-three letters in your genome, uh, in, in the, in DNA, sometimes can cause a disease, and sometimes it doesn't do anything.

Actually, many times it doesn't do anything. But there are specific, specific areas in your genome that if you have a different variant or a different letter there, it can cause disease. And in, in many cases, uh, because the combinations of, of these changes in your, in, in the genome, over three billion letters, right, is so vast that for clinicians and, and scientists, we actually know only a very small portion of which variants are causal to disease or not.

And so there's these benchmarks from, um, folks who collect, you know, variants, uh, different hospital systems and, and clinics, um, and some are known and some are, are, are still, you know, unknown. Folks have created some benchmarks from ClinVar or TRACIM to basically, the idea is, given a, a mutation in a DNA, can you tell if it's gonna cause a disease or not, right?

And so this is a very good setup for DNA models because, um, they're probabilistic. And so when you do make a change, it basically can modify its confidence or pro-- the probability of predicting a next letter. Uh, and in this case, we've sort of leveraged that predictability of these models. Um, they've essentially seen and been trained on so much DNA, in particular human DNA, they kind of understand what's common in a sense, and usually common or conserved, uh, across different other folks, usually typically means healthier.

And so if it's less common, you can think of it this way, if it's less common, the model can pick it up and, and, and sort of predict that it's potentially pathogenic or, or disease-causing. And so we, we've taken some of these, um, benchmarks, and when you introduce a variant or a mutation, there's different types of mutations in, in variants.

So sometimes you can delete a, a letter altogether, you can just flip it, um, you can remove big portions of the DNA, uh, but these are generally single variants. Um, and in, in this case, this is where, um, previous DNA models really struggled, uh, especially on humans. Um, and so we, we showcase that not only is it, um, you know, competitive or capable for humans, but in many cases, it's state-of-the-art, and, and actually m-most of these cases, it's state-of-the-art.

Um, and I think the exciting part is that areas where other models were, you know, currently on the frontier, they still were lagging behind quite a bit in terms of where in the genome. So in the genome, there's coding and non-coding regions, like protein areas, uh, protein regions. Meaning areas that are actually the-- coding the structure of a protein versus other areas that do other things like regulate what genes are expressed. Yeah. And so, um, these non-coding regions are largely regulatory.

They kind of control how much or when to use a certain gene or turn them on. Um, and in many cases, these non-coding regions, these regulatory regions, variants there, you know, mutations there, um, are much harder to predict if they cause disease or not. And I think, uh, what's exciting about the new generational models we're building with Omni is that that's where we shine, especially.

The models are able to pick up mutations and be able to distinguish if it's disease-causing in these non-coding and, and especially long-range areas. Um, so I think this is particularly exciting to a lot of geneticists that have struggled to use traditional bioinformatic tools or statistical methods because they've, they've largely focused on coding regions, which is only about, you know, one and a half, two percent of the genome.

And, um, turns out many, if not most of the diseases are in these non-coding regions. And so there's been a real desire to build models that can actually pick up these variants of d-- um, of, of disease p-- causing variants in these non-coding regions. I have several questions to this. Yeah. First, um, just while we're here, you know, for the, for the listeners, um, there's this column o-on this, a benchmark chart called, uh- Borzoi, reference number four in the blog post.

I think for maybe some historical context, could you talk about what this column represents? And, you know, maybe this also help give context for, you know, the omni column at the, on the right. Yeah. Yeah. Yeah. Great point. So what we show in this benchmark here is really t-taking some of the representative models or the strongest models, um, in, in the deep learning side, and also in the traditional methods.

So we have Evo2 is the latest previous genomic model that our, our team had worked on. And then Borzoi is-- it's also a DNA model, but a very different kind. Essentially, it's a, a supervised model that s- predicts from DNA, um, functional genomic tracks. So it, it, it, it too, is inherently multimodal, but it's not, it's not a language model, so it doesn't predict like a next token prediction.

It goes from a DNA sequence directly to a functional genomic track like chromatin accessibility or gene expression. Um- And these tracks have been annotated extensively by- Yes... people writing their dissertations and whatever. Yeah, yeah. [laughs] Right. So the, the big difference there is that it's a supervised task, right? So it requires labeled outputs.

And in, and in our case, these language models, they do not require labeled outputs, right? So you're r- you're doing raw pre-training on un-annotated sequences. So, you know, why is that desirable? Well, there's a lot more data, a lot more genomic data that's not annotated. Actually, most of it, pretty much, pretty-- [laughs] in many ways, all of it, almost all of it is not annotated.

And so being able to learn from an unsupervised manner, hugely desirable, right? Mm-hmm. Um, for, uh, us, we wanted to showcase the benefit of pre-training on, on raw genomic sequences and, you know, comparing it to state-of-the-art, uh, models in other, in other spaces in DNA. Yeah. And so your prediction is, when you say it's unsupervised, how does the unsupervised property work?

Like, like, how do you convert the output of whatever your model is to an actionable, like, ranker score, whatever? What I talked about before is, uh, you know, to train the mo- omni model, it's pre-trained, so it's unsupervised, but when you're pointing it at a specific task, there, there is a, there is a supervised step.

So it's just, you know- Mm-hmm... taking a smaller data set that is labeled, but essentially what we're doing is using the likelihood scores. So the raw outputs of the language model, which basically you can think of it like a probability for predicting what the next letter is. We can essentially showcase the probability score, the likelihood score for, um, the mutation versus the wild type, so that's seen in the reference genome versus the mutation in this particular case.

And then we'll, we'll have two scores, and then you can think of it as like using a, uh, a ratio of the two to showcase basically how different are you from-- how different is this mutation from, um, from a normal or baseline, basically. Mm-hmm. And then that's, you can think of it as like a surprise factor, um, that the model is able to use and leverage, and then we can use that to essentially, um, score, uh, an actual prediction for disease or not.

Mm-hmm. If that makes sense. Yeah. So, uh, omni auto-regressive, is that... or is it diffusion or something else that you can't tell me? [laughs] Uh- Yeah. We- Okay... we, this time we, we're not describing the exact makeup, but- Okay... um, Evo was auto-regressive, and I think it was the first large scale auto-regressive. Mm-hmm. And so I think for us, we don't tie ourselves down to a specific training objective. We use every tool in the box, toolbox, essentially. Mm-hmm. Yeah. Okay.

But for this specific benchmark, you're going along and you're just using the, the likelihood distribution of, you know, the tokens and some tokens are, you know, the model thinks these are unlikely, and that is probably because some evolutionary constraint, like this doesn't show up often, and because it doesn't show often across genomes, it is probably going to cause problems and, you know, people will not survive, so on.

So you think that is, uh, basically your, your ranking metric or something? Yeah. It's, it's one interpretation- Yeah... of, of what, you know- Mm-hmm... how the model is thinking about it. And, you know, very similar to in, in, in natural language, you can describe the same kind of paradigm. And there is additional case in mid and post-training to- Mm-hmm...

leverage m-more than that, I suppose, 'cause we can teach it specific structure and, and benchmarks- Mm-hmm... so that, uh, it can build on top of, of what you just described, which is like what's common in nature, but also because it's a specific task for disease, um, variant prediction, then the model has additional, um, training introduced during mid-training to showcase and to, to, uh, add additional learning power, essentially.

What are some examples of that? It's a little bit big, but- Again- Yeah... probably secret sauce to some extent, but can you give just a gist of what that looks like? What are the kinds of things you would throw in there? We, we, we would actually throw in the score to itself. So, um, you know, that, that, like, that I mentioned, um, a ratio. It's a ratio of wild type versus mutation.

I-- That-- I would say that's more of a, a zero-shot method where you don't even have to do any mid-training, and that's what Evo2 is doing in particular in this column. So Evo2 is, it's not fine-tuned, essentially. There's a EVE column, which people are basically fine-tuning, you know, using the scores from Evo2, um, that's from Goodfire.

And that's also, you know, folks that we, we, we greatly respect and, and they kind of showcase that these models are, are able to be state-of-the-art when you can, when you can fine-tune them as well, not just zero-shot. And then our model is introducing sort of a step above that, not just fine-tuning, but also introducing structure into...

By structure, I mean the, the format of these benchmarks into the model itself, that, that Q&A style formatting during mid-training, which is what gives us an extra boost even. I see. And extra boost, but I think the other benefit too, um, that, you know, we didn't emphasize too much in the blog, but I think it's really convenient for practical use for scientists, is that you don't-- You're, you're doing this without taking the embeddings and then tapping on a head and then, you know, doing some like regression, which- Is the extra step, it's extra hurdle.

Can you imagine ChatGPT, if like y-every time you ask a question, you had to like fine-tune it for a certain domain? Um, we've done it so that the model is flexible during mid-training to be trained on many tasks at once. And so the-- that fine-tuning, that, that last step of training the embeddings doesn't have to be done.

It's out of the box at that point. You just prompt in a certain format, and it will, um, you, you know, the certain format tells it which task you're gonna do, and then will output the answers in that, in the desired format, basically. How careful were you in dev-- designing this post-training scheme to avoid kinda data leakage with the ClinVar, CreateGM, RNA-GM, and so on- Oh, yeah... like these datasets?

Like, how, like, how confident are you that there is no data leakage either, like accidental or something upstream and that, you know, because I've-- I would not be surprised if a lot of these sequences showed up also in your training data- Yeah... even in a, you know, unsupervised sort of way. Yeah. The short answer is we're extremely cognizant of, of the risk of data leakage and, and ex-extrive- Mm-hmm... extremely hard to, to not mislead or, you know- Yeah...

be careful. And so we, yeah, we have bioinformaticians that are able to basically comb through the data and curate, dedupe, and align sequences to make sure that things that are similar potentially to what's in the benchmark are not there. Mm-hmm. And if they are there, we, we remove it. And so, yeah, so we, we actually have, um, steps to, to, um, QC the data quite extensively. Yeah. Cool. Yeah.

Um, I guess maybe before we move on, I, I think it, it's really cool seeing that there are these supervised methods which previously several of these numbers were, let's say, within the error bars, if not just straight up beating what came before them, and- Yeah... now you have significantly improved upon that. [chuckles] Yeah, we're super excited. Yeah.

And I, I think for the longest time, there's this area or this other method called CAD, which has been state-of-the-art, and state-of-the-art for a reason, which is what the [chuckles] sort of by design, they'll take the best methods and kind of do an ensemble, right? So they'll take, they'll take up another-- even if the best method is another previous model, um, they'll mix it with like an SVM and just like throw the kitchen sink at it.

And so you, you can see why it would be the best, right? And so that was the bar for us. We're like- Mm-hmm... if they're gonna cook the kitchen sink at it, like we're not gonna, you know, cherry-pick one model and say we're better than that. We need to beat what's possible, humanly possible now, like across everything.

And so, um, yeah, our, our, our researchers are, were, were, uh, setting their si-their sights on that [chuckles] to, uh, to see if they can actually improve performance across every method. I'd be interested to see there, there's a some discussion of chain of thought, and that- Yeah... that broke my brain a little bit when I was first, [chuckles] first hearing about that.

I'd be really interested to hear about what that even means. Yeah. Um, I, I had to pore through the blog post to really understand that. Yeah. Yeah. I think this is a really just a, a, a taste of where we think, um, the design capabilities can move toward and, and way more usable for folks. Um, so chain of thought, um, stems from natural language community.

I, I believe Jason Wei at OpenAI, um, showcased the first examples. And, and really what the breakthrough there was, um, showcasing that these language models perform better when you just show your work, essentially. You show the steps of how you came to a conclusion or an argument. And it turns out e-even if the, they were like simple steps, but it just gave the model a chance, [chuckles] you know, may-maybe it's sort of like philosophic, who, who knows exactly why it works. But, um, essentially f-feeding more tokens in a-and giving it more scratch space to think.

And so people just think this is the, um, sort of the, the beginning of reasoning, uh, for these language models, this ability to kind of, um, get to an answer by thinking to itself, by itself. And so it seemed quite successful in language, very successful, and that's why you have, um, a lot of agents that just spend just tons of tokens, right? Just [chuckles] show-showing its work, right?

Um, in many ways, it stems from this chain of thought paradigm. In, in biology, we saw very little of that. We saw-- We started to see some of that in, um, protein design a li-a little. Um, and so we wanted to push that and, and showcase that, you know, well, at first explore is that possible in DNA? Like, what does that even mean in genomics?

Because you don't really have words that describe, you know, its thinking. Um, so how do, how do you take that para-- that same paradigm and, um, introduce it to a DNA language model? And so what we did was a simpler version in so many ways. Uh, we had this dataset of RNA aptamers, so just think of it as these desired sequences with some kind of fitness score associated with them.

So we took this large dataset that had RNA input and a fitness score. The fitness score go high, it's good, right? The simplest version. It's a big dataset, and so what we wanted to showcase was that if we show the model progressively better RNAs in a, in a series of steps with its score, right? So you have like low scores first, and then you gradually move up the chain. Can the model continue that trajectory on its own?

And then, you know, in the final step, w-- does it self-optimize to a point where it's like the best score it can get? That was the experiment. Can, can we do that? And so we took a dataset, a large dataset of aptamers. We held out a portion of, like, the best-performing ones, and we show-showed it only the, the lower ones, but then we ranked it, right?

So we showcased lower scores with the RNA aptamers, and then it progressively got higher, and then c-- asked the model to just, like, continue with that pattern. And it turns out it was able to recapitulate some of those higher scores that we had not shown it yet. Um, and we're actually in the process of validating them at wet lab right now, so we, we didn't get to show it here, but we, we want to know, right? Actually, can it not just do this in silico, which it can.

It showcased that it was able to continue this trajectory and create Design plausible aptamers with higher fitness scores. And now, uh, we think this is, uh, you know, obviously, if this works in the lab, we think this is a hugely, hugely valuable paradigm that can be pr-pr-- that can be pretty much applied to [chuckles] every other type of biological sequence.

That's what usually what you have. You have sequence, so you have some kind of fitness score or desired output. And if we can get models to eventually learn that structure and, you know, basically just show a series of progressively s-stronger sequences, the model can then predict the rest. That's a very powerful paradigm.

Honestly, a bit surprised about this, um, that specifically this task saw strong improvement. Maybe my personal bias is coming in here, but RNA is, uh, somewhat notorious for not having good co-evolution data in terms of, like, constraining structures, right? J-- For, you know, viral genomes, um, oftentimes there is strong evolutionary pressure, but for mammalian, usually RNA does not have evolutionary pressure. And- Yeah...

I think the community has, um, seen that very, very clearly the predominant pressure is, like, RNA will code, um, you know, will, will carry information, coding information. Yeah. So I, I, I, I mean, I'm wondering, like, genomes carry lots of different information. They code for proteins, they have regulatory elements, and, you know, different types of genomes have different types of structure.

So I, I'm wondering, where do you think the-- this capability might have emerged in this language model? Oof, that's a good question. And, and honestly, I, I'm not sure. Like- Mm-hmm... we, we're, we're surprised too. One, because the model is pre-trained on, on DNA- Mm-hmm... and it's really just, like, mid-trained, um, on RNA very, very, you know- Mm-hmm... in a, in a small way.

Yeah, I think what your, your intuition about the DNA having a lot of evolutionary effects or, or, or information is probably where. And so I think this is hinting at the idea of why we think it's so important to pre-train on, on DNA genomes and genomes first, um, and then sort of add additional modalities on top- Mm-hmm...

because you get a lot of transfer, and you want-- y- and the modality generalization is the thing that we're working toward. You know- Mm-hmm... what I didn't talk about as a company, for the company as well, is this idea of building toward general biological intelligence, where we are unifying a lot of the different so-called, like, languages or modalities of biology.

At the end of the day, they stem from DNA, and I think people have not exploited th-that fact as much. It's usually s-really specialized, uh, domain-specific model or modality-specific models and, and not leveraging a lot of inherent shared structure from other modalities. And so one example of that is, like, you know, when people talk about virtual cells, for ex-- this is a little bit tangent, but, um, you know, they, they tend to focus on just RNA and, you know, you know, transcripts, a-and they want to generalize to describing an entire cell. But a s-- obviously, a cell is a lot more things than that.

In my mind, if you want to learn a system, you want to learn from all the signals or sensors of that world or that system. Um, you know, if it's a cell, um, you wanna fuse, you wanna understand the DNA, you wanna understand the metabolomics, the epigenomics, the proteomics, and that, that's when you get closer to, like, quote-unquote, "a virtual cell."

And in our, our minds, we don't even wanna stop at just the cell, but we wanna, we want to fuse all of these sensors across all of biology. Will it get us to a, a, a super intelligence that understands every component of minutia of biology? Who knows? But I am confident that this type of paradigm will get us a hell of a lot further than we are now.

Like, that's- [chuckles] Like, that's my bar- Mm-hmm... is, like, can you make something m- far more useful than, than now? Mm-hmm. Right. So I, I'm curious, are you focusing on eukaryotic cells? Are you focusing on, like, h-human genomes? Do you-- have you gone so far as to do viral genomes? I mean, there's a lot of DNA viruses, but it seems, like, plausible that RNA, that there's a lot of RNA sequence, virus sequences out there, and I'm not sure fundamentally they would be much different in terms of training.

Yeah. Um, I'm, you know, I'm curious, like, what's the scope of that, if you can talk about it? Yeah, absolutely. We are interested in all domains of life. Mm. So, uh, here we focused on humans in particular because we thought, um, this was an area that of, of previous models- Yeah... uh, Evo and Evo2 were not as strong and sort of got a lot of, um, feedback from folks asking, you know, "What are these models useful, useful for?"

"Uh, they can't understand human genomics 'cause it's too complex of grammar and rules and DNA. It's too, it's too noisy, it's too-- there are too many repeat characters and all that stuff." So we, we wanna showcase, we think this is actually useful, and it can be applied to humans, and it's sort of the most complex of the complex in s-in some ways.

But, uh, I think there's opportunity to apply these models generally to, to every form of, of life. Um, so we absolutely are interested in prokaryotes and, and viral. Um, viral in particular, we care about especially for biodefense and biosecurity especially. And, um, I think there are also lots of therapeutic applications that we can learn from microbial life of, you know, maybe obviously for some folks.

Uh, in particular, they-- and I mentioned that folks had used Evo to generate, uh, the first AI genome of bacteriophage. Turns out you can use bacteriophages, uh, potentially for AMR or antimicrobial resistance. You know, if you have a superbug bacteria infection, which in the-- around the world is about two million deaths from bacteria infections still, the idea of using, uh, viruses, designed viruses to target specific bacteria has been done for a long time, uh, particularly in Eastern Europe.

And, um, there's a potential to make a new class of antimicrobials that are not like antibiotics, but very similar, that can be used just like it. Um, and so I think we're gravitating toward things that are high impact and the potential to save lives. Um, and so we're We, we don't stop at just one type of genome.

I think we're interested in anything that's beneficial to humans. Maybe going back to my question about, um, Solex and RNA, um, predicting kind of a chain of thought of RNA evolution. I, I'm curious, was this model trained on RNA sequences or sequences which have some, um, might have evolutionary pressure on RNA structure? Only during mid-training. Only during mid-training. Okay.

So, so, so pre-training is all just genomes and DNA. Mm-hmm. And so the only time we introduced RNA was for this specific task where, and only RNA from this dataset. Okay, okay. So, so not even outside. I see. I see. Yeah, yeah. It was-- So this really was something along the-- There is something encoding RNA structure in this model to some degree, maybe. Um- Or either that or implicitly, maybe. Yeah, implicitly.

I would say implicitly. Um, I mean, 'cause- That's cool. Yeah, 'cause I'd, I'd say se- se- you know, sequences, as we know from proteins, like implicitly should learn structure from- Yeah... from the just sequence. Um, yeah, and so we, we don't, we don't add in two D or three D information at this point- Yeah... but we absolutely plan to. [laughs] Yeah.

So, so, um, just so I understand, first of all, the chain of thought idea is this is a demonstration of it, but the idea is that anything that you can get sort of a s- a training set that has a sequentially better, um, measurement of some sort is maybe a, a candidate for this technique. Yeah. And so can you just describe for this particular experiment, just so we can understand how we're mapping chain of thought to this data set and what the data sets have to look like.

Can you just describe how is the... For this, I know this, this wasn't your data set, but, um, how was the data collected in such a way that you could map accurately from sort of fitness or whatever to a particular phase of or part of the data set? Yeah. So I, so I'm less familiar with how the data was ke-- um, actually sent us or generated from the experimental point, viewpoint, but they are validated from a wet lab in the real world when it was collected.

So it has some kind of fitness score, uh, I, I believe through- I, I can maybe provide a bit of context for this- Yeah, please... if you want. I mean, so the idea here is you just generate a bunch of random sequences, and then you take those sequences and, um, so you have like an aptamer structure, which is essentially like a, a switch with RNA, which sort of when something binds to it, it will do something like cleave, um, off a sequence.

And you can use an NGS, like next generation sequencing readout, um, very high throughput. So you create lots of these different sequences. You, I guess, in this case, they were targeting a, a specific HIV, um- Yeah... like protein or a genome or something. And if it binds, you can, you basically get the signal of like, you get more reads of that.

And so the, the more time, the more sequences which are floating around, kind of like the more fitness, the more likely it is to bind. And then you take those, and then you mutate them again, and you kind of iterate on this. Right. So you have this iterative experiment- Yeah... where you're progressively using the Petri dish to basically- Mm-hmm...

identify the most fit- Yeah... um, thing. So, and then you, what you're doing here is you're basically doing this same experiment- Mm-hmm... in silico. And they're actually doing very high throughput. Like, I think there's like ten to the eleven or something sequences. Some like really high number of kind of sequences explored in parallel for this experiment.

It's- But the key is, is the biology of the experiment is actually doing the filtering, right? Yes, yes. Yeah. And so, and so- Yeah... you can imagine other types of experiments where you could apply the same kind of idea, where you're, you're doing this, like progressive refinement of something which is very common in biological wet lab work.

And, and so if you're capturing those intermediate states, then you can maybe feed them into the model and do that kind of thing. Absolutely. Yeah. So another example is, um, for the antimicrobial resistance, I think that's something we're, we're very interested in as well, the ability to selectively target specific bacteria strains or kill bacteria strains.

Yeah, that's measured basically a score of zero to one, uh, of how effectively that is done. And so showing progressively more effective phage genomes, um, and their associated scores for effectiveness fits that paradigm very well. There's another, um, design task that we're working with a national lab to, to do this f- with as well, an, uh, area that is quite different for us, but it's on, um, designing proteins to extract rare earth minerals.

Um, so- [laughs] Yeah. Um, and so it turns out that y- you, you don't just care about, um, uh, proteins that could bind to something, but you want it to be selective. You want it to be bind to one type of rare earth mineral. Um, and so you have scores associated with how much affinity or, you know, binding affinity for each type of mineral.

We wanna essentially do the similar tho-- the similar exercise with, uh, rare earth minerals and show it a series of progressively, you know, desirable scores, not just for binding to this, but like, you know, lower binding scores for other ones, so you can selectively do it. So I think there's the, the creativity in which you can showcase sequence and desi-desirable, desirable sequence with, you know, some kind of fitness or functional output is a relatively intuitive way for people to design, you know, just prompt by, just by prompt engineering, essentially, which I think is, uh, is very exciting to, to, to explore more. Following up on the, uh, rare earth mineral extraction, I, I find that is an interesting use case for this.

In fact, maybe one where you probably wouldn't have a comparative advantage compared to some other techniques. Sure. Because it seems like your strengths are probably in larger scale design across like entire organisms, but focusing on individual proteins, that may be one much more of a, like a structural task.

And I, I'm curious, like going to talking about ClinVar, the argument here is that the model now is a really good statistical representation of what type of mutations are common or not common. Yeah. And I think in order to do design as a s-- If you want to design structures, I think you want to understand structure. If you want to understand disease, I think that- Is I think more natural- Mm-hmm...

or for many diseases, is much more natural in terms of like, a, a population genomic sort of way. Totally. So I, I'm, I'm curiously, where do you think your strongest competitive advantage is, um, you know, in using this strategy, and do you think that your model understands structure in addition to function? Yeah.

I'd say w- what drew us to this particular application, rare earths, and, and our strengths in general- Mm-hmm... w- why we thought it might be suited for it, uh, is two things. One is, I think in, in co- in areas where context matters. Mm-hmm. So yeah, you're right. So a protein design task where you're just designing structure, not, maybe not naturally where we see ourselves, um, you know, competitively advantaged.

But, uh, in the rare earths case, our hypothesis is that context matters. A- and what matters here possibly is, uh, certain microorganisms with proteins that, you know, have the function that we desire, we can potentially prompt and provide us context for this is the neighborhood in which, you know, tell Omni, "This is the neighborhood of genomes or microorganisms that you should search for new proteins."

So it's really a, a mi- like, sort of like a mining exercise, and what we can showcase to our models that other folks can't for like, you know, like protein- Mm-hmm... structure models, is that we can feed in, um, non-coding regions before that gene of interest or that protein of interest, and then ask f- for the model to p- provide variants essentially.

So like, "Mutate this protein, but, you know, know that you're in this microorganism, but show me different variants that you've seen in nature, or combine different things that you've seen in nature given this context." Um, and I think that is one reason why we can make very evolutionary diverse sequences and potentially phages. Folks have used Evo models to do this with toxin, anti-toxins in a sim- very similar mech- uh, technique.

They've prompted on things upstream from the proteins of interest, and then asked the model to kind of generate a bunch of plausible other ones. Um, and I, and I think this case for rare earths, that's super exciting 'cause then we can come up with plausible variants and then test them relatively simply for what they bind to and selectively bind to, which I think is really interesting for this, this, um, this partner who cares, for example, about securing the supply chain of rare earths for the US strategically.

Yeah. And so we thought that's absolutely worth something that, um, is worth supporting. So your point is not just you're designing a protein, but you're designing an organism which generates a protein, and this protein has an action. But or, you know, or is it, it's that, that the, the way to design this is you need to understand how, through the tree of life, interactions of proteins with rare earths have or with certain minerals have occurred.

Yeah. I'd, I'd say in this particular case for, like, the rare earths- Uh-huh... we, we're not really interested in the, in the organism- Mm-hmm... like designing the whole organism, but I think the organism does tell us about what proteins are plausible and- Mm-hmm... the selectivity is not, or the way they evolve, not just ar- around the protein itself, but also the regulatory elements around it. Um, and it helps us narrow and provide additional context.

I guess I think of the genome or DNA broadly as sort of the imprint of the physical world into DNA. And so there's parts of the DNA that we want, but the things around those parts that we want also tell us a little bit about the context of where it came from and how it came to be in its function. And so I just think of it, you know, like g- g- going back to that word, but context, I think context matters, um, in, in many of these applications, or at least in, in some of them.

[laughs] Uh, yeah, I think context matters a lot in biology. Um, I think maybe one of our big bottlenecks is the lack of context and how- Yeah... we as humans mostly approach biology in terms of a very engineering, like let's isolate individual systems and, uh, systems biology approach is very hard to get any sort of meaningful quantitative predictive power.

So maybe going off on a tangent, but I, I mean, I'm curious, you know, going to context and thinking about how context scales to an organism, you're talking about right now two million length context, right? Yeah. You know, I think the human genome is, uh, roughly, you know, a thousand times longer. [laughs] Um, so, uh, or but even like a lot of, you know, say bacterial genomes, if you're trying to engineer them, are, are quite a bit longer than that. So how do you leverage something which has a long but still finite context compared, you know, to, to do synthetic biology across, like, large organisms? Yeah.

And maybe actually maybe for context, how did the Evo two bacteriophage design work, which was probably much more than two million for that as well? Yeah. So I could, I could speak to the, the Evo bacteriophage just a little bit 'cause it's actually a separate group that- Mm. Mm-hmm... worked on that. But, um, that context was actually pretty short, actually. Oh, wow.

The reason why they started with phages- Mm-hmm... is 'cause it's amongst the, the shortest genomes. And so I believe it was something around six thousand base pairs. Oh, okay. Wow. Oh. That's really short. Yeah. Yeah. [laughs] So extremely short. Viruses are insane. They- Yeah... they're incredibly efficient. They impact a lot in there. [laughs] Yeah. Yeah.

Um, but yeah, no, I think your question about how, how do you get longer context with something smaller- Mm-hmm... um, is a very key question that we, I think broadly the AI community is constantly trying to- Mm-hmm... fix and be creative about. And so, I mean, I think that's a large part why we as a company are an AI research lab first, because we, we think the innovation needs to be constantly pushed.

It's not a space where we can just grab open source models and, and expect that, you know, many of the tasks that we care about are just gonna be solved. Uh, we c- we wanna continually push the envelope, and so context is one of the key research areas that we, we drive in. I'd say that's probably how we got our name, 'cause we f- worked on long context before it was a, it was, it was a thing, I guess, like twenty twenty-three, twenty twenty-two. Yeah. Long context.

You, you and your team, I mean, your, your collaborators have a long history of these state, space, uh, state-space models- Yeah... doing, you know, pushing context, um, like what seemed insane at the time. Yeah. I mean, now I think- Routine of books, the big slabs- Yeah, exactly... but at the time was like orders of magnitude longer. I, um, I don't know if you wanna talk about that a bit. Sure.

And also, I think one, maybe one thing which I think is really fascinating is how in, for both this model and other ones, you really go on a first principle way, like diving into the architectures of how GPUs work and designing models which are, you know, exploiting the architecture of GPU in addition to-- I mean, it's not just, oh, we're, we're building longer.

It's like, what can we do special given the compute constraints we have to push the boundary? Yeah, absolutely. Yeah. If you wanna distract our researchers at, at Radical Labs- [laughs]... uh, this is, this is how you nerd snipe them. [laughs] You talk about, you bring up long context and kernels, GPUs, then they're like, "Wait, did someone say kernels?"

[laughs] And they, and they start trying to, to figure out, you know, how to make things fast. Yeah, long context, special place in my heart 'cause that's what I focused on in my PhD at Stanford, and that's how we started thinking about DNA and, you know, going back a little bit for fun. [laughs] We were looking on working on language models in general, and then we noticed our models were good at long context.

And so that's the progression of like how we started working in the space was like, oh, these models seem to be really efficient on long context. Um, and then with Michael Pali, uh, my, my lab mate, he worked on the first design of, uh, Hyena, this convolutional architecture. And then we started thinking like, let's push this further.

Let's see what new c- new applications open up if we really lean into long context. And we asked, "What's the longest sequence out there?" And, and eventually, unsurprisingly, we, we, we landed on DNA. We were like, DNA's gotta be the longest, three billion base pairs and, and, you know. Um, and, um, we started thinking like, okay, what's being done there? Like, what kind of context lengths are people doing there?

They were doing super short. They were doing like a, like one or two thousand base pairs or tokens at a time. It's like a th-- you know, w-way smaller than what one would want for DNA. I mean, most human transcripts are like three K or so. [laughs] So that's not even like as, you know, most-- That's not even what you need to represent, like a protein or- Yeah... most of the time. Yeah.

And so it's clearly a, a need there and overlooked, and we, we, we saw it as a, a way to initially, like let's see if we can do something that we had no idea if it was gonna work and, and no idea who would want it. And so we just started tinkering around, and it turns out out of the box, relatively out of the box, it was, it was doing pretty well at reading DNA. And then it led to, okay, let's-- People keep asking about like DNA.

They didn't really care about our language stuff as much. And they would say like, "Can you, can you do longer? Can... You know, what can you do with it?" They always ask, "What can you do with it?" We, we didn't know, really. Um, and then for some reason, latched on to this idea of like, well, no one's writing DNA. Can we, can we get it to write DNA? And that was this, a really simple question, but in, in hindsight, almost seems obvious, uh, that yeah, you would wanna write DNA and design, design it.

But when I was first pitching the idea of Evo to folks, I spent six months, which I guess in bio world is not that long, but I spent six months going around saying like, "Hey, if I generate DNA, like would you find that useful? Like, what would you do with it? Um, would you back us? Like, would you wanna be a part of this?" And crazy enough, most, almost every scientist at Stanford I talked to thought it was a stupid idea.

It was-- I was like, this would be so cool, like generating DNA. Um, how, how much would we accelerate the field? And then people would say like, "What would you do with it?" [laughs] I'm like, "I don't know." [laughs] And, and then I would get that comment or comments like, "That's not possible. Like, we as humans don't understand the rules. How could you expect an AI to learn it? You can't even tell if it's right or wrong.

Like, you can't tell the AI, yes, that's right or wrong. How can you expect it to learn it?" Or, "There's too many repeat characters," or, "DNA is too noisy of a distribution. Like, there, there's no real rules in there. There's a bunch of junk in there." I heard all the reasons, and I, I was just so stubborn about it. There's gotta be a use case from, from being able to generate DNA.

I just, it just feels right, and I didn't know what it was. So it was, it really was an experiment of like what happens. And so when we first trained Evo, I remember, um, we had no idea if it was gonna work. We're, we, we, we had no idea. We didn't know what thing was gonna emerge. We just thought, let's just train a big one, which is k-kind of ridiculous, but somehow, you know, folks at Arc, they were like, "Sure."

Yeah, why not? Why not? [laughs] Let's see what happens. Let's pay, let's pay some money, uh, for the GPUs and, and let these crazy kids, uh, train a model. And, um, I remember when the first s-result came back and, and it, it kinda like gave us chills of like, oh, maybe there's something going on. Uh, which was somebody took the model checkpoint, the first one, and threw it at, uh, protein gym, one of the protein, um, benchmarks, and turned out to be competitive with pr- uh, protein-specific models.

And we were like, okay, that is pretty surprising because we never ch- told the model what's proteins versus not. And there's actually, it was, you know, there's also protein and DNA, right? So there was one piece, I think it was just surprising that it was actually competitive with protein-specific models.

Um, and this was like the first experience, and then they just kinda kept on coming one after another, like, "Oh, this is competitive here. Oh, it's state-of-the-art on RNA and R-- DNA." And we started seeing like, oh, it can learn across different modalities, like not just DNA. And then, then we felt like, okay, this, there's something there.

And so it kind of went from there, and then, you know, folks at Nvidia were like, "Let's, let's back Evo two. Let's make this even bigger." And then Greg Brockman from OpenAI was like, "I'll take a break from OpenAI and take a s- four-month sabbatical and, like, help these crazy kids out." And, you know, then we're Slack messaging Greg Brockman at three AM trying to de-debug our code, which is wild. [laughs] Yeah.

So it, it just, it, it went-- The trajectory was very surprising in many ways, but at the same time, you still got a lot of feedback like, "Mm, you know, what are these models good for? What are they, you know, what are they useful for in the, in the real world?" And so that's what really motivated us, uh, to start a company. We thought- What we showed was just really just the taste from like an academic flavor.

In a similar way that language models, when they first too, you know, from natural language came out, people asked sim-similar questions like, "What are these things good for? Oh, cool, it can write some jokes for me. Is this gonna lead to like an all, you know, all-encompassing A-AGI that can automate everything?" They did not think that, right?

They thought, "It's like a toy, it's got emerging capabilities," and they would extrapolate the potential. And I, in many ways, I think what we're seeing here, uh, is even more exciting or, or reminiscent of that trajectory for, for DNA. But is Evo two even was showing the bacteriophage, and that, that wasn't a persuasive.

I mean, that's, uh, almost a scary example, right? As you mentioned. So the people didn't see that it-- like that isn't a light bulb moment for people. Yeah. So I mean, and some people, right, I think, um, either fair or, or, you know, rough critiques, you can even s- you know, play dev-devil's advocate and b- and say what it's generating is kind of, you know, pretty close to nature, and you're kind of re-recapitulating just kind of a small variance.

And so there's-- I think there's a lot of ways to critique and, and, and kind of, you know, minimize the, the potential, which can be fair argument, like. So I think at this point, we think what is more important is not just pushing on the science and like, cool, this can be done and like kind of leave it there as a sort of thought experiment.

But like how do we actually use this to improve human understanding of disease, improve, um, treatments, make better rare earth mineral extractors? How do we actually make this useful is what we care about now as a company, in addition to some of these more scientific questions. So maybe that's a good segue. Um, there's-- that brings up two things for me.

One is the McIntyre stuff that's in the end of the blog post, and then also the biosafety stuff. Those are both, I think, important- Yeah... applications. So can we-- let's do McIntyre first. Can you talk a little bit about, um, th-- and this is actually kind of building on the work that Evo did as well, I think, um, in the-- I remember in the Evo paper, there was some McIntyre work in which they were discussing using the mo-- reversing the question and using the model to extract insights about biology, right?

And so, so and this has actually become a common theme with, I think maybe biology more than any other domain, is that people, because it's a scientific question and it's learning patterns about the world, you can actually say, "Okay, well, what patterns did you learn?" Yeah. So I think McIntyre is an emerging, uh, field in, in bio that we're extremely excited about. And admittedly, we are on the early side, I'd say.

Um, so we're building up that team, but so far what we've showcased and, and been excited about is looking, really analyzing the embeddings and some of the activations in the model. Um, and so this idea of McIntyre for, for, for bio i-is borrowing a lot from the natural English community currently. Um, you could think of what the models are doing as compressing a bunch of information it's seen, right?

And so in this compression, it's really basically distilling it down into the key components, key patterns that, uh, helps it understand the data or learns the distribution of the data. And so what we're trying to do is probe the models to see what did the model distill into its weights and its activations or sort of like the outputs.

And so for us, we started off with a lot of the outputs of the model, so the activations, and we, we wanted to see what kind of structure, what kind of visualizations can we see that help us understand, um, some of the complexities of, of DNA, which was a ton, right? And some of th- some of these complexities can, can range around GC content, certain motifs of transcription factors.

Th-there's a bunch of regulatory types of patterns and motifs that, um, the model we believe has to pick up to be able to do its tasks, right? To understand whether a disease is caused by a variant, um, generally, it's going to compress all these different motifs and distill them into the model weights.

And so our job is to then find these and see if we can distill a pattern or structure that can be generalized to other cases where we don't understand the patterns, right? So that's sort of largely the goal. So in this case, GC means the two nucleotides that are, um, the fraction of those- Yeah... in a sequence. Exactly. Yeah. Yeah.

So the frac- the fraction of the G's, G and C letters in a genome, which I guess amongst the more simpler things, but, um, also, you know, simple things as repeats, number of repeats. Um, I think transcription factor or TF motifs is another, um, one that is especially interesting for folks 'cause I-- eventually we're going to get to a case where we can design transcription factors, meaning we'll have certain transcription factor patterns via like ChIP-seq, you know, those other modalities.

Transcription factors are, are... Well, you go ahead. Transcription factors are, uh, molecules that can bind to DNA, and they would alter essentially the gene expression pattern. Um, and, and so it has this regulatory effect that doesn't modify the DNA itself, but can modify sort of the effects of DNA in the products that DNA makes.

Um, it can have a lot of implications on, well, pretty much everything in your body. So it can modify disease states, it can modify, I guess that's a lot of aging related researches around transcription factor, um, design. And so we think being able to understand some of the motifs via, uh, DNA, but also additional modalities will eventually let us be able to design transcription factor patterns as well.

And so I think this is re-very exciting. In, in many ways, the combinatorial space of learning these transcription factors, like what binds and where they bind and what effects it causes, is just far too vast to be able to do this, um, in a manual way. And so we wanna take a data-driven approach- Yeah... to, to learn some of these motifs. The complexity here is partly because the transcription factors themselves, um, coding genes, and so that you can, right, those can regulate each other, and so that you have this, th-that's where that combinatorial effect comes. Yeah.

So y- like just narrating for the listeners only, we're kind of marching through these increasingly complex and higher level Factors all the way from GC content. We started, now we're looking at disease, which is maybe the most complex thing you could or you, you are, uh, pointing towards in this analysis. Yeah.

So I think, um, what we-- our first idea for previewing this was, and, and what we're working toward just broadly, is, um, this idea of mapping the manifold of disease, right? So manifold, there's like the sort of representation space of what disease looks like to a model, [chuckles] in terms of the output scores or embeddings.

We believe that there's a lot more structure that can be gleaned from understanding some of these outputs. Um, and so, you know, mapping this manifold or the landscape of what the structure of disease, and obviously there's many diseases, and so, um, I think would be a big exciting area of research for us, um, to, to actually drive motivation for mech-int-repin bio.

I think, I think we're just really scratching the surface because if you can get this much of-- glean this much of insight, potentially from just DNA, which is, in my mind, just one of the sensors that you want to ultimately fuse into, you know, modeling biology, then being able to do something similar across all modalities is something like-- all, by all modalities, I mean protein, RNA, epigenomics, you know, the attack, uh, chromatin accessibility and, and, and methylation patterns.

All these other different types of sort of phe-- molecular phenotypes around DNA just presents such a huge opportunity that just ki- is kind of laid in front of us that is all green space, green greenfield. No, I, I haven't seen anybody do, um, this level of sophisticated techniques from, you know, machine learning, deep learning into what I think is gonna be the most important, impactful area of understanding and applications for AI.

That's what we're excited about. And I think we're just showing a preview, a ve-very simple preview from the DNA-only models. Um, but in our next generation of models, which will be increasingly multimodal, we're talking dozens, it's a very exciting moment for us to showcase. Sidebar on that. Is language, nat-natural language, one of the modes?

Not yet. Not yet, okay. Yeah, but it will be. So I'd say there's a lot of questions about how to fuse that with, with bio, but I actually think it's pretty re-- will be relatively straightforward. I think the more tricky, the tricky part for us is actually how to fuse the, um, biological signals more so. Fusing language, we've se-- there's a lot of examples with that with, like, with image and, and video space.

Um, so we f- we feel pretty good about that, and we've done some early experiments with language, and I think that will make it extra accessible for folks when you can connect it to language. But I think, um, this, the part, that recipe that folks still are trying to figure out is how to do this across, uh, biological modalities. It just seems like a natural way to be able to do chain of thought, right?

Exactly. Yeah. Absolutely. And I think, um, you know, especially when you start having, um, chain of thoughts from orchestrating tools, things like cloud science, I think the idea of incorporating language, um, it's, it's already on a lot of people's minds. If you don't have any questions, uh, I-- let's talk about the, the biosecurity. Biosecurity. For, for us, we as a company thought it was very important to have a, a dual mandate, wh-what we call a dual mandate.

It's this idea of essentially being cognizant and, and feeling responsible or wanting to feel responsible for the capabilities that we're enabling on the design side. So if we're going to create models that can design function into sequences, um, we believe and see a gap in companies being able to safeguard that technology and make sure that it's used responsibly, um, increasingly more.

I think folks have-- I was just at a panel last night, uh, a panel on AI scientists, agents that can do scientific discovery. And one of the last questions was, what's-- what are some of the biggest risks or doomsday scenarios with AI can learn-- learning about science? And every one of them talked about, um, biological weapons.

And, and at the same time, I'm cur-- I was curious because I was like, "Okay, so then what are any of these folks doing about that?" And basically, I, I didn't hear anything about that. And I-- you know, these companies, I won't say their names, but-- and I look at the companies, they don't have big efforts in those spaces.

So anyways, we felt it was important as a lab, in a lab, that a team that was both building the design capabilities is actually best, also best suited for building the defense capabilities, because they're basically the same models. Yeah. A model that is good at generating, turns out is also very good at discriminating or predicting with, uh, you know, if a sequence is pathogenic or not.

So we, we felt it, it was not just, you know, from a principle standpoint necessary to, to work on both biosecurity and the design, but that it was, you know, strategically, it just made sense, um, as well. And so we felt this resonated with our team, but also the broader community, um, folks in, in, in the US government and abroad even, um, that this-- there was a, a clear gap and, and need for a type of entity to exist and, uh, to do this.

And so, yeah, we, we felt it was necessary to build it into our mission. And so for us, what we have done and plan to do, when you look at it, or I should say biodefense, biosecurity sort of has three or four different pillars of, you know, strategy for biodefense. The first one is around detection and surveillance, right? You know, broadly, can you detect from the environment if, uh, a sequence has a pathogen in it, something that can cause disease?

Let's say from, you know, swabs at a airport, a nasal swab, or, um, just a sewage system, right? Collecting samples. The next is attribution, which is once you've detected a danger, can you figure out where it came from? Is it natural? Is it from, you know, a random country abroad? Is it engineered by a human from a specific lab in a country? And that helps you figure out, you know, what, what to do about it, right?

So this next idea is around countermeasures. So once you've detected, figure out where it's from What do you do about it? Can you make a counter agent? Can you make an antiviral or antimicrobial? And then fourth one's generally around, um, deterrence, but that's more of like a, a government kind of level thing. But yeah, so we focus m- primarily on the first three.

So we create tools that can, given a sequence, detect if it's pathogenic, but also what we felt was missing from the community was not just detect, you know, broadly if it's pathogenic, but characterize the heck out of it, meaning, uh, what parts of the sequence are dangerous, what genes, uh, for example, attributing where it came from, being able to not just look at its, you know, sequence and match it to a database, but be able to attribute its signatures, and then also a big component to make this effective in the first place.

The community, the, the biodefense community broadly focuses on sequence matching. So they'll take a sequence, and they'll basically l- align it to a known database and see... Say, "Have I seen this before? Does it match, you know, this list of known pathogens?" But I think what's emerging a-and a concern for a lot of labs is, well, one, new stuff, right, if it's not on your list, and two, things that were intentionally obfuscated to not be detected in sequence space, meaning the letters matching up exactly, but also function space.

Yeah. Right? Because basically, models that we're enabling now, they will be able to be function aware or structure aware. And so that means for concretely, you can have a sequence that has the same function, like a pathogen, but actually look different in terms of the letters. And you can imagine based...

Actually, what, what we have on the screen here still is the smack interp thing, and there's this sort of manifold that the, that the model constructs internally that, that is kind of coding for these, among other things, functions. And so that you can imagine how it would be able to say, "Oh, well, that's, you know, maybe genetically quite different or at least somewhat different, but still has a similar function." Exactly. Yeah. Exactly. So I...

What we talk about in our, um, defense blog is this idea of things that can function similarly. They can start having separation in terms of, um, what the sequence looks like while maintaining the same functionality, right? So this is can happen in nature sort of, um, naturally, but also what these AI models allow you to do is in- also intentionally do that as well. So have the same function or functional capability, but have diverse letters, basically, diverse spelling, but describe the same thing basically. And so in this case, this work from Microsoft called paraphrases, so it's like, you know, kind of rewording things- Mm-hmm...

where they showcase that you can, for example, use protein language models to essentially keep the same structure, which structure implies similar function, but then change the spelling, right? And not just that, they wanted to test that if you have this capability and you send this through existing detection systems, would it break the system? Like, would it actually detect it or not?

And, um, I think one of the f- interesting facts that maybe the general public, but ev- most biologists know, is that there's these DNA synthesis companies, right? Where you could basically send a design of sequences and get back a DNA s- you know, DNA molecule like it's an Amazon package. Like you just send it off, and they'll send you the physical DNA of the design.

Um, they'll, they'll manufacture it for you. And this is-- this runs the scientific community, right? It's the, the pipeline that allows people to do research and understand biology and make drugs and everything. So it's, it's prevalent and, and it's, it's public. And so I think a con- one of the concerns for folks is, well, and one of the concerns for us when we first started w- working on this was in a world, for example, where agents are prolific online, presumably billions and trillions of agents building and taking all sorts of actions on the internet, it's wild that they don't have any tools to basically tell if it's making anything dangerous or not.

And so that was, that was literally our first motivation of like, we should probably make something that can detect if something is dangerous or not, right? Um, so- So you guys are a filter in the- Mm-hmm... for these manufacturers that they can say, "What am I making here? Is this dangerous?" or whatever. Yeah.

And maybe you need a-- you, you could have an exception if you were like some licensed lab or something, and I'm doing something dangerous. I know I'm doing it. Please let me do it anyway or something. All sorts of cases. So that's, that's one scenario. To be fair, the s- many of the DNA synthesis companies have detection tools, but I would, I would, uh, strongly hypothesize that they're not AI based and, and fairly...

They're probably not robust. They're pattern matching. Yeah. They're, they're all pattern matching mostly. Sorry. Um- Yeah, please, please. Maybe this, maybe this... I should ask this question later. I, I'm just curious about, you know, we've all... everyone who is working in bio has tried to use Fable and instantly gotten picked up on literally everything you type in. [laughs] My website, for example. Yeah, yeah.

[laughs] You, you, you know, you can't do anything in Fable without, um... But in terms of people like scientists exploring synthetic biology and creating new sequences and designing new sequences, it seems like it, it'd be very hard to get, you know, to, to have a, an ROC curve which you can t- you can live on that doesn't like impede novel scientific research, which legitimate purpose is.

How do you avoid, you know, even with an F one of point, you know, nine nine, which you're not even close to right now, [laughs] I think, you know, that, that still could easily, if there are billions of sequences, you know, sequence generated a day or at least like a year, I mean, I, I think you could really have a lot of...

It seems like a very hard balance to- Yeah. Yeah. It's, it's a tough question, right? Mm-hmm. So I think broadly, the way we look at it is, you know, if we thought about like, how do we one hundred percent stop [laughs] the dangerous design, I think it's a l- it's, it's a harder question to ask. I think the question we asked is, on the capabilities design side- There's plenty of folks pushing the frontier of that.

When we look at the defense side, do we see frontier technology being applied there? And the, the answer to that was no, right? So we, we, we saw this huge gap, and we wanted to sort of aid, come to its defense, I said. [chuckles] Come, come to its, um, uh, its aid to give it a boost, right? Yeah. So I think that perspective, it's, it's an easy choice for us to say, "Let's push... Let's bring that to a better, you know, head-to-head match against the design side." Is it gonna solve everything? Well, I think that's what we're gonna aspire to be, but, but, uh, realistically, there's always gonna be cases where it can get, get around, right? And I think that's one r- that's a big motivation for why we think, you know, safety for language models and chatbots is one, one layer, right?

But also, you're right, there may be things that kinda just get out there, get past that anyways. And so what do you do about those cases? It's, it's already past, you know, the chatbots, right? It's already, it's already aided someone into making dangerous sequences, so it's out there. Um, I think the cool thing about our tools is that we, what we're building is tools for the folks that care about things that, that's already out there, out there in the environment, that's made it somewhere, and now there's this whole ecosystem that we wanna build into that does the surveillance, that does the attribution, that does the countermeasures.

We wanna boost that community, right, and build stronger tools for that space. Does it have to be perfect to be useful there? I don't think so. I think we can, we can be helpful and, and move the biodefense community forward in terms of, you know, bringing AI technology and AI frontier technology to their aid without being perfect in it.

So that's kind of how we look at it. So maybe another way to say, so threat model, the-- it's not even clear what threat model you're actually trying to defend against at this point. But the capabilities don't currently exist at all. So, you know, you provide something, and now this can be worked on with, uh, in terms of a larger regulatory framework or, um, government or, you know, nonprofit, whatever larger framework, and this now provides you a tool that the community can build upon.

Yeah. Um, even if it's not perfect, it's, it provides a starting point, and if you don't have a tool, then you can't do anything. Yeah. The bar is the improvement, right? So the bar is, like, where are you at now? Can we move the needle? Yeah. Can we move beyond sequence-based matching, alignment-based matching? Absolutely. I think there's tons, and I think, um, the interest is, is only growing.

I think we're hearing a lot of chatter at the, at the, uh, regulatory side, you know, from different politicians and different folks in, in think tanks. It does seem like folks are mobilizing. So we're optimistic that this gets out and is more top of mind for folks to actually take action as opposed to just talking about it.

'Cause right now it does feel like there's a lot of talk, and then one of the reasons why we felt like let's build tools and, like, put it out and give access to people, 'cause there's been a lot of talk about AI companies saying, like, "We should do bio defense." But then like, okay, what does that mean? Like, what are you gonna do about it? [chuckles] And so that was our, our approach. So if I can just steelman this for a minute, though.

Can, can you go to the other diagram where you have the... Yeah. And if you click on them, then they show other similar- Mm-hmm... similar compounds or similar, um, genomes. So f- arguably just steelmanning the, the opposing viewpoint, if you are on the frontier, which it seems like you are, and you certainly strive to be, you're moving the frontier of the attack and the defense at the same time.

[chuckles] Yeah. So, like, right now the, what you're doing clearly helps, but that, that maybe if you're on the frontier, then that doesn't really matter for, in the long term. So how do you think about that? I mean, y- I guess you just have to nerf what people are doing or something. Yeah. The mindset of w- was hoping to get folks to start th-thinking about it as less of a, "Oh, let's make a biodefense tool," and, like, call it a day. It's basically an arms race, right? It, it...

Similar to the cybersecurity community, you're gonna make better technology, especially with something like Fable, that can potentially attack, and that means you're gonna have this back and forth. The design side's gonna get c- more capable. The defensive side needs to t-try its best to get ahead, right? And then that, then that just motivates other folks to do... pass that too, right, uh, on the design side.

So I think inherently there is this arms race style dynamic that, um, the way it appears to us is that the defensive side has been far, far lagging. And so what we wanna do is bring the defensive side closer to, to par, essentially. So that, that's the mindset I, I picture or the framing I think about it. The cybersecurity, uh, analogy is interesting, but I think it differs in some, like, key points.

So first of all, I think that with cybersecurity, with a sufficiently strong model, you might actually be able to close all loopholes which are not sociological. There are certain ones which will always be hard, always be ways of getting around things. But you could, in principle, catch every single exploit, and I think that might be possible in the future.

And then you can patch them, and as long as people stay updated, you know, you're secure. We have fixed genomes, right? So you can't patch a human. [chuckles] I mean- No... uh, so once something... So I think that in some sense the, the, you know, attack surfaces and/or the, the way that you're defending against that is much higher or much harder.

But then maybe the, the converse is that the, uh, it seems much less likely that someone would have the incentive to go on offense to the same degree, and also the, the, like, barrier to entry to succe- even if you can print out arbitrary DNA and the, the process of going from that to making a successful, you know, virus, especially one which doesn't kill the person designing it, is actually quite large.

So- Yeah... I, I guess I'm-- maybe I, I'm curious, like, what is the single... From your opinion, what is, like, the single biggest threat that we actually have? What would the thing which keeps you asleep or keeps you awake at night? Is there something in particular, or is this, like, you just think this is something we need to build, and let's build it?

It-- [chuckles] One, it's hard to get in the mindset of a person wanting to design a bioweapon. [chuckles] Yeah. Um, so we're, we're, we're, we're not trying to, um, necessarily think of all the potential people or scenarios that, you know, you know, bad actors might, um, work on. What I think broadly, what I worry about is we're lowering the bar for how much expertise is needed and the speed at which folks can engineer these kind of things, and that means the volume's gonna just exponentially increase at, at some point.

And so there's a mix of, yes, there's intentional worries because there's certainly state actors that have had biological war, uh, programs, very, very, very large ones. [chuckles] Mm-hmm. And, and one of our advisors on a company has, you know, physically seen these facilities and, and de-decommissioned them.

And so we've heard a lot of stories about, um, states actually being motivated to, to create such weapons. Um, that is one concern, and w- I think during peacetime it's less scary. During wartime, it's particularly scary. I think the unintentional ones are also things that, um, in my mind, potentially more likely in the near term, where folks do try to generate things and control function and inadvertently things that, um, maybe, maybe they tried to make a, a s- a certain thing to understand and to study, but it got out, right?

'Cause these things can be hard to contaminate-- uh, to, to, to contain, for example, and there's just a, a leakage. So I think those scenarios potentially seem the most likely. Does it keep me up at night? Not necessarily. I'm, I'm much more of an overall an optimist. I think building this type of technology is ultimately a game that you weigh out the pros and cons and the cost and benefits.

I think the benefits far outweigh the potential harm, and so that's why we, we work on it, because ultimately we do think it's gonna be an engine for, for discovery and human health improvement. But at the same time, we just felt like the, the defensive side was sort of losing its, th-this, this arms race, and so we work on it as well and try to push the frontier.

But I, I'd say overall, I think as a community s-of, of researchers, but also folks in the, you know, the policy side, I think the, the community is resilient enough and, and has the ability to, to mobilize to get ahead of it, and so I'm, I'm very optimistic about it. We have two questions- Sure... that we, we like to ask, um, every guest. So the first one is, if you could, by fiat, remove a bottleneck that is important to you, what would that be? Interesting.

Um, could I give two answers on this one? Sure. [chuckles] I think the first one is, um, kind of a cop-out 'cause every AI lab says this, like- GPUs... GPUs. [laughing] GPUs. We- Yeah, we've done that answer a few times. We, we could use GPUs. Um, the other one, which I think is a little more philosophical, I think in this space, in what we're trying to do, aspire to do, i-is reinvent how scientists do their work in this space, and I think one hurdle we run into is folks who, and you see this in, uh, many domains, folks who are, have-- the more expertise you have in something, the more pessimistic you become about that space.

Mm. And I think you especially see this in bio, where you know a disease area or modality so well, and then someone introduces something else new, and you're like, "Oh, but what about this and this and that?" And they're very pessimistic, and probably rightfully so. But also, I think what I've noticed at the company, what we strive to do, and the folks that we try to bring in are domain experts that do know that field, but also are still dreamers, meaning they still do have that imagination and desire to change how things are done.

I think that barrier, we see that a lot in the field. I think if we embrace that more, we can see a lot more progress and step change that I would love to, to see. Brings us to the last question- Yeah. Great... which is, uh, yeah, is there something that you want the audience to take away? A single message. Yeah. I think, um, one message is, I think folks have had, especially into the AI research community, felt like there was a choice they had to make sometimes to either work on the frontier of AI technology and that was, like, consumer-related apps or enterprise-related apps and, like, they just had to work on chatbots.

And that's, that's, that's the cutting edge technology. But also do, you know, I think when people think about the true potential of what AI can do, and I think a lot of it is about improving human health, understanding our biology. But people have felt like they have had, had to choose and like, "If I work in that space, I can't work in the frontier of AI." And what I would like people to take away is that you don't have to choose.

I think we can work on the things you truly care about, that you think will push humanity and work on cutting edge technology, and that's what we're trying to build at, at Radical Numerics. Yeah. And I mean, clearly, and I encourage people to read the blog posts, the architecture, the mech inter, there's a lot of innovation that is going into building these models, and I firmly believe that biology is really on the forefront of AI.

Awesome. Yeah. Glad you feel that way. Cool. Amazing. Cool. Thanks for chatting. Yeah. Yeah. Thank you for- That was fun... taking, t- making the long trip, like [laughing] blocks away. My plea- anytime. [laughing] Thank you. Appreciate the invite. I had a blast. Great. Thank you. Yeah. [outro jingle]

🔬 Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin