The Evolution Revolution: Scouting Frontiers in AI for Biology with Brian Hie

Hello and welcome to The Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week we'll explore their revolutionary ideas, and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my co-host, Erik Thornberg. Hello and welcome back to The Cognitive Revolution. Today I'm speaking with Brian He, an assistant professor at Stanford and innovation investigator at ARK Institute, who is at the forefront of applying AI to biology. We begin with a discussion of biology's grand challenges, understanding the causal web of interactions in biological systems and designing interventions which are both effective and narrowly targeted. Brian offers his perspective on our progress in these areas and insights into how AI is already beginning to change the game. From there, we discuss 3 of Bryan's recent papers, each representing a different facet of his work. We start with Mechanistic Design and Scaling of Hybrid Architectures, a paper that describes a semi-automated way to build novel machine learning architectures from primitives including the attention and state-space mechanisms. This paper shows that small model performance on toy problems, aka microskills, is highly predictive of their performance at much larger scale. And while this paper is very much focused on machine learning techniques, the insights that they derive are useful for the long-sequence challenges of biology. Moreover, the techniques themselves suggest that a sort of directed evolution applied to machine learning is likely to create a Cambrian explosion of AI architectures. From there, we move on to EVO, a hybrid attention and state-space language model that demonstrates surprising emergent capabilities understanding higher-level biological concepts despite being trained solely on DNA sequences. We discussed the model's ability to identify which genes are critical to an organism's survival and also to generate novel CRISPR variants, as well as the precautions that Brian and collaborators took when developing this model and Brian's overall perspective on the biosecurity landscape. Finally, we discussed Brian's work on using AI to guide the evolution of antibody complexes. In another result that demonstrates AI's remarkable ability to generalize out of distribution, Brian and his team have been able to use a model that was trained with 3-dimensional structure data representing single proteins to create antibodies that can be up to 25 times better at binding to their targets than their natural counterparts, with obvious implications for drug discovery and development. Along the way, we touch on broader themes in AI for biology, including the challenges of interpreting models that are trained on data types which humans did not create and really only partially understand, The compute requirements for cutting-edge research in AI for biology, and the need to account for evolutionary dynamism in AI system design. As always, if you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. And if you have any questions or feedback, feel free to reach out either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. Now, I hope you enjoy this survey of important frontiers in the application of AI to biology with Professor Brian He.

Brian He, assistant professor at Stanford and innovation investigator at ARK Institute. Welcome to The Cognitive Revolution.

It's great to be here. Thanks, Nathan.

I'm excited for this. You have been a part of some really remarkable research publications recently, a couple that I kind of felt as I encountered them, came from different parts of the world, but you were involved with both. So I think it's going to be a really interesting conversation to understand those projects more deeply and also to try to make some connections between them. I wanted to maybe just start off with really kind of big picture survey of the landscape of AI for science and maybe more specifically AI for biology. I'll just kind of float my understanding at you and you can critique it, tell me what I'm missing or how you would characterize things differently. And I'm not a biologist, so I'm probably going to get some stuff wrong.

I'm not a biologist.

But I see kind of 2, at least 2, but 2 that really jump out to me, grand challenges in biology.

Yeah.

One is understanding how things work. Like what is the sort of causal web of interactions between all the different parts that make up a living being, what interacts with what, what's upstream and what's downstream. And ultimately, for the purposes of what most people are really interested in, which is like better medicines and longer, healthier lifespans, figuring out all those interactions is necessary to understanding what we want to target if we want to intervene. in biological systems, most importantly ourselves. Second one I identify is once we have a target identified, then we have to figure out how can we actually intervene and do something with respect to that target in a way that is both hopefully effective but also specific, so we're not causing other sorts of unintended impacts on the system. My sense is that that second one has been We've made major progress on that in the last few years, and AI has certainly had a major role in that and only see an increasing role for that going forward. The first one seems like we're starting to see some signs of figuring that out, and some of your work is, I think, a hint of maybe what's to come there, but it seems like we're maybe not as far along. How would you react to my big picture overview?

No, I think you actually Like have a pretty good understanding of the current big picture in biology. So yeah, especially when it comes to say like drug discovery and curing disease or preventing disease, I feel like that the first question, what is the actual causal mechanism of the disease? That is probably the most important question or the most pressing one and also the most complicated one. I mean, the reason why most drugs fail in the clinic is not because you failed to develop a good binder for your target or you didn't have the right small molecule or antibody for the target. The reason why most drugs fail in the clinic is because you just went after the wrong target to begin with. Or you have these endless debates like in Alzheimer's, is amyloid the right thing to target or not? Or in some diseases, we just have no idea what causes the disease. And so it's hard to design a binder to something when you don't even know. What to bind to. So I think definitely that's a big question. We're interested in this question. From a machine learning perspective, I think a lot of this question is fundamentally data limited, actually, at least in my view. Because in terms of actually profiling biology and in disease, we can collect lots of observational data, actually. So you take a snapshot at a given point in time, And you can profile a ton. You can look at the whole genome sequence, you can do a bunch of single-cell RNA sequencing. So you count up how often those genes are actually turned into RNA, and if the genes are turned into more RNA than others, like maybe that gene is more relevant for some given condition or not. We can collect a lot of this observational data and get a pretty full snapshot of the biology at 2 different time points or in 2 different conditions. But biology is so complicated that even if you have really high-resolution snapshots of 2 different conditions, and you can just observe this, you know, a lot of things that are different between the 2 snapshots in health and disease, most of those are just totally not related to the actual causal mechanism at all. Actually, it's because biology is so complicated that there's a lot of just spurious correlations. And oftentimes, The actual causal factor will just not be measured, or that the measurement looks like nothing has changed between the 2 conditions, but actually that's the real causal target. So this is what we deal with in biology, and especially in machine learning, the only way fundamentally to get to this question is to actually collect a lot more interventional data. So like you actually make, or perturbation biologists like to call it, or it's actually instead of observational data, you need to collect the interventional data. And hopefully that will get us closer to this question.

Yeah, keep going. And this might be a good time to bring in a little bit of the Scanorama work that you've been involved with.

Yeah.

Because it seems like that is a step in this direction, right?

Yeah. So early on in my graduate career, actually, which the Scanorama paper is part of, I was working in this area, like this single-cell transcriptomics. So the idea is, you know, you can measure which genes are turned on or off via the RNA sequencing across millions of cells. And then the Scanorama paper in particular was asking, you know, you have a bunch of different— you profile the same tissue but across different laboratories, and they're all actually like slightly different. So can you try to, you know, align these 2 datasets together, find which cells are actually the same across the 2 different labs, and kind of integrate information across multiple— data sources. And yeah, single cell at the time was, and it still continues to be a big area of effort for algorithm development and now machine learning modeling development. But that was my first exposure to this kind of grand challenge in biology. Like, how do you separate the causal changes from the observational changes? And we did try, you know, in that line of work, Scanorama, but also other algorithms. that I was working on for this problem. Fundamentally, I was just looking at the data and being like, there's only so much causal information you can learn from observational data alone. And so that's why I kind of took a hiatus from working in this area. But I think there are a lot of efforts now in trying to collect this causal data. And once these efforts have completed, I think it's then probably a good time to revisit this question. Yeah.

Yeah, it seems like we may make some in silico progress as well. That seems like another extremely promising thing, but let's build up to that over the course of a couple papers.

Yeah. There's many ways to do it. You can collect the data. I mean, if you could go the other way, bottom up, like from the level of like atoms to molecules to interactions and get perfect mechanistic model of systems biology and model everything in silico just from first principles upwards, then this is another way of potentially getting at this problem. Maybe without doing a ton of interventional data, you have an actual mechanistic theory of the cell or something that describes the biology perfectly. And then you can also think about causality. And I think we've been interested in actually both directions and maybe we're working on both directions and maybe they'll meet in the middle or something.

Yeah, that meeting in the middle, that sounds right to me. So let's go through a couple of your recent papers in some depth, and then we can kind of flesh out this future vision, and it'll be much more grounded for folks having understood what you've already put out in a series of recent papers. So I guess maybe just one quick sidestep, like, how do you think of yourself as a scientist or researcher today? Because we're going to go through one that's like purely ML, then we're going to go through one that's like AI meets bio, and then the third one is like the most on the— Still AI and bio, but like more toward the bio side. What's the taxonomy of activities in your own mind?

Yeah, I feel kind of, it's weird. I get asked this somewhat by people, like you just described, like, you know, someone's a biochemist or someone is a mathematician or someone's, yeah, I think I still say computational biologist, but in biology, especially for a long time, computational biology I feel like was kind of looked down as being like not as rigorous, or maybe the models were not as good and the biologists kind of didn't believe the models. And so rightly, like kind of looked down at computational biology. But I think computational biology as a field is actually really getting a lot better.

Rapidly ascendant for sure.

AlphaFold was great for this. But I think like there's one version of computational biology where you're like in the mushy middle and don't know anything about both. But I think there's a better version of computational biology. Where it's actually hard and it requires you to become a decent expert at just fundamental biology and machine learning for the algorithm side and the intersection. And so I think that's the version of computational biology that I try to practice. And so it does end up looking like sometimes we publish things in ML venues and sometimes we publish things in journals and stuff.

Yeah. Okay, cool. Well, let's start with this paper, Mechanistic Design and Scaling of Hybrid Architectures. This is a pretty much pure machine learning piece of work, though, in a way, if I really squint at it, I can sort of see it as like a beginning of sort of ecology of ML architectures. That's maybe a bit of a stretch, but honestly, that is something that I do have a sense that we are entering. a new era where the last few years obviously has been just totally dominated by transformers and attention is all you need and so on and so forth. And now what this paper is looking at is saying, hey, we've got now a couple of different core mechanisms. Like, of course, attention is still part of what you need, but there's also new stuff. Like, we have the state space mechanism that is working remarkably well too. And this paper looks at how these different core mechanisms perform differently on what I like to call microskills. And these are things like in-context recall, fuzzy recall, noisy recall, selective copying, compression, memorisation. These are things that are not like your grand big picture large language model most impressive examples, but they are the microskills that make those larger things possible. Or at least that's how I think about it. And so now that we have this states-based model to kind of compare and contrast against attention, and I think, by the way, more probably coming not too far behind, we've got to figure out what are the recombinant, if you will, forms of this that are ultimately going to work best for us. And this is what this paper sets out to investigate. So with that, Prompt. Tell us about the investigation that you did there and what you guys found.

Yeah, I think it's quite interesting. And this paper was really driven by a really talented, I think at the time, graduate student, or he was maybe working together, Michael Pauly. And it's interesting that you brought up the need for taxonomizing or something. And I think Michael's first version of the code, it was called Zoology or something. There's now a Cambrian explosion of different architectures. So yeah, actually in biology, in terms of speciation events and diversification, often what we see from the fossil record is that, you know, there's these long periods where nothing happens and then these bursts of evolutionary diversification, the Cambrian explosion being one of them. It does actually seem like we're going through something like this now in terms of architectures where people are revisiting different architectures for various reasons. And actually, in terms of like— The evaluation and testing harness for comparing different architectures to each other is also formalizing. Rather than someone develops a new architecture and it's not controlled in any way, or the set of evaluations for this new architecture are different from the set of evaluations for the other architecture. And maybe this has hindered architecture design for a long time, and I think is one of the things that this than the MAD paper is trying to address, actually. So actually being able to formalize, making sure that you design the experiments well and that they're controlled, making sure your scaling laws analysis are FLOPS controlled between, or actually controlling for the amount of compute, to even higher level things like the set of the primitives or the micro skills, actually developing a set of tasks that You can use to evaluate different architectures and compare them with each other. And I think that will also be useful for this kind of Cambrian explosion of architecture design. Another interesting aspect of the paper that actually did seem very biological to me is that a lot of the motivation for these microscales or primitives is that there may be more efficient correlates of scaling laws. So if it does better on some set of tasks, then that will also translate to better scale. But doing a very extensive scaling laws experiment, like Chinchilla style, like FLOPS control and everything, is actually pretty compute intensive, especially if you're in an academic lab or something. But actually finding good low resource correlates of these scaling laws, I think, is also pretty interesting. And we come across this problem a lot of times. As well in just biology, right? So, you have some really resource-intensive readout, like a clinical trial, which takes 10 years or something, or even going into a mouse model or a non-human primate model. But, you want— part of a big thing to accelerate any kind of development process is a good cheap correlate of the more expensive thing. So, does it work just in a Petri dish or does it work in a chemical reaction? format? And does that correlate well with success in the animal? And does the success in the animal correlate well with success in the clinical trial? So I think that aspect of the paper too felt very biological to me. So yeah, I think that's my reaction to that prompt that you gave.

Hey, we'll continue our interview in a moment after a word from our sponsors. Yeah, so just to flesh out a couple details, you mentioned the term MAD, that is Mechanistic Architecture Design. So part of the approach here is to say we have these kind of building blocks, we can sort of manipulate them with code, we can shuffle them around and try all sorts of permutations. And that doesn't have to be a handcrafted, let's say, artisan sort of undertaking. It can be the sort of thing that you just grind through with a higher-level code framework to develop and test all these things. And then I think you said this pretty clearly, but it does bear reemphasising that the paper also shows that the performance at very small scale on these microtasks does in fact predict the performance of the architectures at much larger scale. Obviously not going all the way to frontier model scale, but enough orders of magnitude there that you can start to be somewhat confident that you could probably extrapolate it out a bit further. And also, maybe the most important thing is then you encourage or you sort of create reason to believe that further research at very small scale might have similar predictive properties. Obviously, that's not always going to hold, but it does seem like a really important just foundation laying that I expect a lot of people will probably build on top of. Do you have microskills that you think are kind of most important to the sciences? And do you see any like core weaknesses in microskills right now that we don't have architectures?

You mean for the actual, this is not for, you know, new architecture development, right? This is for the practice of doing science, you mean?

Yeah. Well, partly where I'm going with this is I've become quite fascinated with the KAN architecture recently. That's K-A-N from Ziming Liu in Max Tegmark's group.

Yeah.

And, you know, they're trying to do AI for science. The way that they present their results is often about greater interpretability and kind of composability. But I sort of see it as likely to become another one of these core mechanisms that if there is a MAD-2 to come, then this might be another one of the core mechanisms that gets folded into the shuffle. And ultimately we see like what share of layers it can earn for itself. But I guess to try to develop my intuition for that, I'm wondering, at some point we may have enough of these core building blocks, core mechanisms. I don't think we're there yet, but my best intuition for sort of how to know if we had enough would be, are there still micro skills that we see the current architectures just can't handle? Yeah.

This is interesting. And I guess I didn't touch upon the main finding of that paper, right? Is that rather than a pure architecture, all the layers are transformer layers or something, or all the layers are hyena, that actually hybridization or just mixing different layers tends to work better than pure attention. And now that 1 in 8 to 1 in 10 layers is of attention combined with Mamba or Hyena or something does seem to be pretty good in practice. Actually, in a model that we have, the Evo model, we actually use 3 out of 32 or something layers of attention. So around 1 in 10 or so. Yeah. I think that in terms of the composition of the layers, as well as the primitives with which to assess the right composition of the layers, I guess it brings up a meta question of how do you actually find the right tasks to begin with? And so yeah, I'm not sure. Yeah, I think there's just a lot of work to be done. I don't want to speculate too much on this.

Yeah, the canned thing, I think, is going to be— it feels to me like there is something. The Arc AGI Challenge has kind of captivated at least a portion of the field recently. And there's this sense that, which I think is a little bit overstated, But nevertheless, definitely has some truth to it that language models can't reason. I always say to that, well, I think they can reason, just not necessarily super well, keeping in mind that a couple of years ago they couldn't reason at all. So going from none to some, to me, suggests that there's probably more to come. And this is where I'm not entirely sure, do we need a new core mechanism, or maybe just keep scaling what we have and that could be enough? When I look at the KAN thing, I do see something really seemingly fundamentally new there. Just the ability to compose functions, the ability to initialise with particular functions in place. You can almost build a calculator into a neural network. And there may be also just logical operators that need to be kind of woven in somehow too. I was just reading the latest Wolfram piece. I think it's probably— I don't expect you've got to that one yet because it just came out in the last 3 days. He, as always, kind of reformulates— in this case, he reformulates machine learning in terms of cellular automata and whatnot. And there's a pretty extensive section on how to create just these very discrete logical operators in a sort of machine learning type context. How do you make a AND gate or a NOT gate or an XOR gate or whatever. Right. And it's like these sorts of things that we know how to make to some extent are kind of being approximated or being converged on in neural networks. But it seems like maybe we could get a lot farther, a lot faster if we like initialized with some of those structures baked in. I don't know if it's necessary, but it feels, because you might just get, again, scale might solve everything, but it feels like it is.

Sure. Sure.

A really nice shortcut and that it will also probably hybridize really well with, in the way that this paper explores.

Yeah, definitely. I mean, this is a big question, right? How much of these architectural changes are actually useful or how much of it washes out with scale? And I think a lot of this stuff will be empirical and we just need to design the right benchmarks to actually test. And just ground the actual model development in actual empirical performance. But I do think in terms of all these, this new architecture development, at the very least, you get comparable perplexity or something, or comparable modeling performance. Just like, it just fits better in our academic compute instance or something. Or the MFU is just higher. The actual utilization is higher. to train the model and actually conferring a kind of FLOPS-controlled just modeling improvement potentially as well. But yeah, we'll just have to see. Though what I like about this kind of era of machine learning is the emphasis on objective benchmarks and just empirical performance. Does this model just work? And I think that kind of mentality for new architecture design will drive a lot of progress. Yeah, GANs are very cool. compositionality and everything.

So before we move on, there were a number of other interesting nuggets in this paper. I definitely do recommend it as one that people should check out. It's quite a thorough study.

You should have Michael on the podcast.

Yeah, I would be interested to do that for sure. I mean, his name does come up across quite a few very cool pieces, but there's tidbits on mixture of experts, on the size of heads on finding that increasing data on a sort of, in terms of within a compute budget, right? Pushing more on data versus pushing more on parameters seems to be more helpful. That last one I didn't have an intuition for. That seems to be kind of the general trend, like LLaMA overtraining, whatever. But I didn't quite have a sense for why that would be better. I don't know if you had one, but I'm curious if you do and any other final notes on that paper before we move on to Evo.

Yeah, I think, no, that is an interesting observation. Yeah. And there seems to have been a general trend to just smaller models, more data, or just smaller in parameters and more data.

Yeah.

I think, for example, like Llama 3 compared to the Chinchilla Optimal or whatever, they go like way beyond in terms of tokens actually consumed by the model, for example. I think In terms of the intuition, it does seem like the models are capable of learning a lot more than we thought they could, or that many thought they could. The models potentially, like, even at 7B or something, they have enough parameters to actually learn a lot, even with lots of tokens. In terms of an intuition for it, I mean, there are lots of parameters in the model and they can interact with each other with lots of combinatorial ways and actually There could be, you know, the same neuron is involved in many different concepts or something. And so there's actually a lot of potential capacity in the models beyond just the parameter count, right? And this has been empirically shown in smaller models, right? Smaller models can actually— the same neuron can be involved in many different concepts or something. So there's like this kind of superposition or whatever. And also like training a smaller model on lots of data is really helpful for inference, right? So we try to think about this too, like when we're making models for biology, a lot of biologists definitely like want to use the model We don't want to release some trillion-parameter model or 100 billion-parameter model and no biologist would be able to run that on their local hardware or something. It would probably struggle to. Even at 7B, that might be very big for someone in the biological community to use. We actually do think about the benefit in inference with just having a smaller model that's just smarter because it's seen a lot of tokens. So I think that's a useful, like you have the resource-rich people maybe spending more FLOPS to get a model that's then helpful for people who don't have enough compute to do inference. I don't know. These are the 2 things that this prompt kind of spurs to mind.

Hey, we'll continue our interview in a moment after a word from our sponsors.

Cool.

All right. Well, let's move on to the second paper. This is maybe the paper that I mention most on the podcast. Which is kind of surprising because it's a bit outside of like my standard fare, but I personally find some really powerful lessons in it. So this is Evo. Full paper title is Sequence Modeling and Design from Molecular to Genome Scale with Evo. And first thing to note is this is using the striped hyena architecture, right? So it's one of these hybrids as we kind of alluded to. Much more state space with some attention sprinkled in. It is long context, which obviously the state space models really help enable, and it's trained purely on DNA sequence data. And yet some interesting things start to come out of that. And the proposition gets floated, is DNA all you need? Probably not, but it certainly can take you farther than you might naively expect. What I take from this, and I want to make sure I'm not overinterpreting it, but what I see here is a really interesting data point on, first of all, prelude to a lot of what's to come in AI for science and AI for biology. But even just in terms of understanding what is happening in these models right now, There's so often a lot of confusion when it's like, okay, it's a language model, we trained it on language, maybe it learned this concept, but it was only— you used to hear people say, well, it doesn't really know, it's just predicting the next token, it's just stochastic parrot, whatever. That, I think, has largely been put to bed now. We can see that there are pretty clear concepts that can be isolated within models. They're not just doing pure naive correlation. But then the next kind of level of scepticism is like, well, they may have learned that concept, but they learned that concept from us. We figured out the concept, and then they're learning that from our representation of those concepts in language. And when I get into that path of discussion, then I'm like, but you need to study Evo because it seems to me that this is a clear case of a model that was just trained on next token prediction, in this case, in the DNA domain actually still learning these higher-order concepts in a way where it didn't get it from us. It's clearly getting it from the raw data. So how do you tell that story? We can get into the gene essentiality and some of the other ways that you get at them.

Yeah, there's so much in that. I think that's actually a very interesting dimension that I actually hadn't thought of before. So yeah, we get the question, is DNA all you need or whatever? I mean, obviously in Biology 101, the question On the first day of class, is it nature versus nurture, right? And it's like, is everything encoded by the genes or is it just purely a product of your environment or is it both? So obviously, like, the environment is important and we don't model the environment explicitly with EVO, but actually we have something similar, like, in language, right? It's like, you know, is this thing that's learning language actually a world model? Or there's these questions about grounding. Like, The world actually does leave its imprint on the actual distribution of sequences that you find on the internet or whatever. And in the same way, the environment leaves a very clear imprint on the distribution of DNA sequences that we sample through evolution and through selection. So yeah, this is, I think, the reason why the models we think are working well and will continue to work well with scale is that The world does leave an imprint on the DNA sequences. So even though we don't explicitly model the environment, there's potentially a lot more about the environment or about, you know, higher level functions and concepts that you can learn from, from just raw DNA sequence. And the advantage of raw DNA sequence is that it's very abundant, like it's really cheap. People have been doing whole genome sequencing, or they couple all their biological experiments to some sequencing readout cuz it's so high throughput and so cheap. And also it's just the most fundamental modality in biology, like all biological information that needs to be transmitted through DNA basically. Or, or, you know, the vast, vast, vast, vast majority. And epigenetic inheritance is still somewhat controversial, but especially for microbes and for most information across generations like that needs to be transmitted through DNA. And then in terms of like for Evo, yeah, and what's nice about biology is that human-level performance in biology is already abysmal. Like, it's not hard to get a model that exceeds human-level performance. And there's so much in nature that nature has produced that far transcends any kind of human imagination. I think there's still a ton of biological discoveries to be made. And organisms, essentially in organisms that we see every day, in our own, in human biology, actually, there's a lot of human biology that's unknown, let alone biology of organisms from deep sea vents or something. Like, there's probably a lot of interesting molecular interactions going on there. And I think trying to have a model learn high-level concepts in a fairly unbiased way across like the diversity of life, like that is what we're trying to do from EVO. And actually, this would be a model that's like definitely learning from a different like data generating function, right?

Yeah.

It's like, it's learning from Evolution, hence the name of the model. And hopefully, yeah, the idea is that if we can find a way to reliably extract these concepts from the model, this is like probably the main challenge. Like the model has learned something. How do we now start to interpret this model and connect it to, to actual biological discoveries? But if we learn how to do that well, we can actually go beyond human knowledge, right? We can actually go beyond, like we can use the model to do biological discovery. In potentially pretty powerful ways. But yeah, I think that's a cool dimension, like as an example of a large model that can go beyond the concepts that humans have generated. Yeah.

So tell in terms of like how this is already useful. I think that I always remember first the gene essentiality thing, but I don't actually know a lot about gene essentiality. So kind of give us the problem setup there and then what this model was able to do and like why that's Remarkable.

Yeah, the problem setup, but the actual gene essentiality problem is an organism has some set of genes and each gene will typically encode for a protein. Sometimes you have genes encoding for RNAs, but each gene typically accomplishes some function and the genes and maybe many genes work together to accomplish a function. But some genes are essential for just the viability of the organism. So you don't have the gene, you're just dead. But other genes, you take that gene out of the system, The organism actually seems fine. And that gene was maybe not even that important or is definitely not essential to the reproductive viability of the organism. And this is useful for a number of things. I mean, it's useful for basic biology. Maybe the more essential genes are older in evolution, but it's also useful for drug discovery, right? So, if you have some very bad bacteria, And you want to design a drug against this bacteria. You're not going to target an unessential gene because if you try to target this unessential gene, it's going to do nothing. You want to actually target the essential genes, make sure that those essential genes aren't accomplishing their function. And by removing those essential genes, you have stopped this bacteria from being alive or pathogenic or something. So a very interesting experiment that we ran with EVO, because it can fit long context, up to 131 kilobases or 131,000 nucleotides. And some genomes are actually shorter than that. So the genomes of these organisms called bacteriophages, they fit into the context of evo, say it's 40 kb genome, a 40 kilobase genome. We did this interesting experiment where for each gene in the genome, we insert a very bad mutation. In this case, it's a mutation to like just at the beginning of the gene that tells the cellular machinery to not actually translate this particular gene into protein. So if you made this mutation into an actual organism, that gene would just not accomplish anything. It would just not turn into protein and therefore not be able to function. Biologists call this a knockout, and you often knock out a gene by this strategy. Like, you insert this premature stop signal telling the protein manufacturing machinery to stop prematurely, and so you don't produce the The whole valid protein. So we make this really bad mutation, which basically translates into a 3-letter insertion into, you know, a 60,000-character sequence. And we make this small insertion into all the genes in the genome. And so for each insertion, we then score it just with the log-probs of EVO. So we just look at the likelihood of the sequence under the model with And without the insertion into each gene. And what we actually see is when we make this insertion into non-essential genes, that likelihood doesn't change very much. It's almost like the model's saying, yeah, sure, like you can put this very bad mutation in these unessential genes. It's totally fine. The likelihood remains high of the full genome sequence. But an interesting thing happens when we put this bad mutation into essential genes. There, the likelihood drops a lot. So it's almost like the model is saying, you know, this mutation in this gene looks really bad. It has low likelihood under the model compared to the unmutated sequence. And we actually can then use this and show— and we can compare this to experimental data. So for a lot of bacteria, we actually— people have done these experiments. They're pretty painstaking experiments. We have to mutate every single gene and see how that affects the viability of this bacteria or this bacteriophage. So there's databases of these experiments and we just compare, like, so EVO has never seen the outcomes of these experiments, right? It's just learning kind of, we think of it as like a density model over all the genomes in the dataset or all the prokaryotic genomes. But even so, like EVO's change in likelihoods is able to predict which genes are or are not essential using these experimental measurements as ground truth. So I think this is surprising in a number of ways. One, the model is sensitive to really small changes in a really big sequence. So, you know, a 3-letter insertion into a sequence, the model can still pick up information from that. We can also show that adding sequence context helps the model distinguish essential from unessential genes. So if we just give it the sequence of the gene, the model can kind of predict this well. If we give it more of the genes surrounding it, the model appears to get information about this gene in its genomic context, and that is informative for Distinguishing whether or not this gene is important. So we can show that the model is actually using information from multiple genes or just beyond the single gene context. It's also interesting because people trying to connect a single or a small 3-letter change in the genome to the fitness of an entire organism, which is really complex, like this is the genotype to phenotype question. This has been a very challenging problem in biology for a long time. a long time. And these are just zero-shot predictions out of the model. So it's not like EVO has completely predicted gene essentiality across all organisms, though in some organisms, it's getting pretty close, actually. So for one type of phage, people did this very painstaking study where they used CRISPR to actually mutate all the genes in this phage genome and determine which ones were or were not essential. And with EVO, instead of doing this painstaking study, we just run the neural network through a few forward passes and basically get the same information out on gene essentiality. So, I think it provides a blueprint for a model that could connect how changes in the genotype, changes in the DNA sequence actually affect the whole organism fitness.

So, if I wanted to, for example, create a universal antibiotic and thinking back to our first grand challenge of like, we don't really have a great sense of what causes what in most of biology. It seems like what I would want to do is take all the known bacteria sequences. Maybe I'd want to limit myself to ones that actually cause disease. Maybe I wouldn't. I'm not sure.

Yeah.

And then, I would want to run this process through and I would want to look for genes that are consistently essential across all of my targets, right? And then I would identify all the things that fit that description. And then I would probably move over to the wet lab and start to think, okay, which of these can I actually hit with a drug and which ones actually prove out to matter? That's like, if that's obvious, unless that's totally mistaken, I assume that that is where this is going.

Yeah, because that's one of many potential applications of— so yeah, if you wanted to create a broad spectrum antibiotic, you want to hit these very conserved genes that are common across multiple bacteria. And also because they're common, they're also essential across multiple bacteria or something. So if you want a broad spectrum antibiotic, you want to target these genes that are conserved or common throughout evolution. And what Evo does, actually, the way I think of it, is kind of like a conservation score on steroids or something. Like it's literally looked at, it's been doing this next token prediction. Across thousands, tens of thousands of bacterial genomes and millions of phage genomes. And so it kind of has a sense of which genes in the genome you don't want to touch. Like these genes, any kind of mutation to them is bad news for the organism, or any change to this gene just causes the organism to be completely selected out. So we don't even see it in the training dataset versus this other set of genes. It's mutating a lot in evolution. It's probably not essential to the organism. So yeah, EVO can help you, for example, pinpoint which genes you don't want to touch if your purpose is to keep the bacteria alive, but maybe you do want to touch those genes if your purpose is to develop an antibiotic.

Yeah, I guess if I were to put my like stochastic parrot skeptic hat on for a second, one thing that you just said there is like, it's a conservation score on steroids. I don't know. I mean, I do, I know the basic concept that genes that are highly conserved must be important. You know, those that are not so conserved tend to be less important. How much does this add on to that understanding? Is there a way to quantify or somehow characterize how much more we get out of this versus, you know, correlation analysis and pure statistical parrotry?

We also get the same sort of, you know, critique or skepticism or just observation from people in biology. Like, these models are just memorizing the dataset and they're purely doing correlations or co-occurrences across the sequences. I don't know. See, this is like, just to like briefly hint on, or like how you would test this, or I think we've been looking especially at regions of the genome that are not conserved in sequence, but that are conserved or that are related to each other in terms of some higher-order function. So there's parts of the genome that are just extremely variable in sequence, like they're actually enriched for de novo or new genes. So these are regions of prokaryotic genomes that have lots of new gene birth. So actually, if you look at— but they all accomplish somewhat of the same higher-order function, actually. So they're often involved in say, the bacterial immune system. So the bacteria needs to constantly defend itself against invading bacterial viruses. And if you have a really conserved gene that is responsible for bacterial immunity, a virus is just going to take advantage of that and hit exactly like an antibiotic, right? Like the virus is going to take advantage of this. So bacteria do not want conserved memorized sequences in these regions. They want to diversify. them in sequence, but they want to preserve this higher-order concept that this gene is involved in immunity. Some interesting work that we've been doing in the lab actually is seeing if the model can also generate things that, in terms of sequence space, look nothing like anything in the training dataset. In fact, these genes look nothing like anything in nature in terms of their sequences. But do they still have this higher-order concept Or higher order function that they're involved in immunity. And it, and, and, and this is like where we're going. Hopefully, like the, the data looks promising, but I think this is how you would, you would do it. And, and I think the models are learning higher order properties. They're not just parroting out trivial, like, you know, variations of the same sequence basically. I, I, I do think that they are capable of, you know, diversifying in sequence space quite a bit, uh, to the point where, you know, this sequence looks nothing like anything that's in any of our existing knowledge bases.

Yeah.

But it still has this higher-order concept of biological function. So I think we have to do the work to show that this is the case, because in the absence of any data, one could most conservatively say that the model is just learning these sequence correlations or these co-occurrence patterns and no high-level concepts. And the bar is high to prove that it's doing more than this, but I do think the model is doing more. Good.

That's great. Would you say that the bottom line kind of state of knowledge today, it is Suggestive but not definitively proven. It sounds like you're in the process of doing that. You believe it, but there's not full knockdown irrefutable evidence in the public yet.

I think we and probably others just need to do the work to show that this is not the case, but I'm very optimistic. And this, I think, will parallel results in language too.

Yeah. I mean, it's going to be very interesting to see what interpretability techniques applied to this looks like. The sparse autoencoder paradigm applied to EVO-2 trained on 15 trillion tokens or whatever, it seems like that is going to be an absolute goldmine of discovery of higher-order concepts, right?

I hope so. Yeah, we want to do this. I mean, we're thinking of various ways of doing interpretability, one of which is you try to train an SAE and you actually do some kind of expansion so that concepts are not entangled or whatever and are associated with Or that there's like some clean thing that's associated with some concept. I think this could be actually pretty cool. The hope is that, you know, this kind of interpretive mechanistic interpretability or this kind of work, well, A, it would hopefully indicate or provide additional evidence that the models are learning higher-level concepts. B, it would enable us to basically do classification out of these models for free. Train, or you don't have to do any fine-tuning potentially, right? You just like have some kind of classifier for some interesting biological system by inspecting the activation of neuron or like a concept or something. And the third is we're hoping that this also lets us do this more sample efficiently. So where you don't have a lot of examples of some system, but you still hope that the model has encountered enough examples of the system that are just unknown that it has learned this concept. This happens a lot in biology where someone will discover some new system like CRISPR. So CRISPR-Cas, there's maybe a handful of different CRISPR-Cas systems that have been discovered. But potentially, you know, in nature, we've just not discovered a lot of different CRISPR-Cas systems or things that are like semantically related to CRISPR-Cas systems that we would want to discover. But we only have what's observed now. But if we can get a concept that's specific to our set of 20 CRISPR-Cas systems right now, then maybe we can use this to do next-generation biological discovery or design, right? So then you find the CRISPR-Cas neuron or the gene editing neuron or whatever, and then you just clamp its activations to be really high, and then you just mine the outputs of the model. So yeah, I think this kind of stuff could be really Really cool in bio.

And you're already kind of approaching that in this paper, right? I mean, the other headline-making finding was that the EVO model already can generate CRISPR variants. This is where I'm a little bit out of my depth on the biology side. I'll ask a really naive question first, which is, why do we need more CRISPR variants? Like, give us a little sense of that. And then Now we have this DNA language model that can generate tons of them, but presumably we have the same kinds of problems that we often have with especially not super scaled language models, which is like they hallucinate. So I imagine some of these CRISPR-like things that it outputs don't actually work. So yeah, give us kind of the rundown of the CRISPR dimension of this.

Yeah, I think that's another interesting component of the study. The study is so massive, like, but I really like this component.

Yes.

Why do we need more CRISPRs? Well, okay. So, I mean, one actually kind of troll answer is that people like to do patent busting and people will actually patent some CRISPRs and then they'll put this like sequence identity cutoff. You know, we patent everything around 80% sequence identity to this molecule. And then—

It's a great answer, honestly.

Yeah. You diversify it with a generative model and then you just patent bust it. But also, when people are designing new tools, because biology is so complicated, there's just some chance that a variant of, say, CRISPR from nature will actually have more favorable properties in the highly unnatural systems that we humans put these systems into. So we actually use CRISPR systems in a very different context from the environment. And so, just maybe by chance, all these things in the environment have the same function or the same properties, but in an unnatural system, just by chance, some will be more favorable than others. And, the idea for a long time is if you have a working system in human cells, for example, so CRISPR, if it works in human cells, great, but it has low activity. Let's just look into nature. Let's just actually look for similar systems in nature. Test them in human cells, and by chance, you know, one of these just might have like a lot better activity in human cells. This typically has just looked like mining existing sequences from nature. So you just actually literally take the sequences that have been sampled by evolution and then test them. With the generative model, you don't need to go out and sequence bacteria from deep sea vents or the sewers or something to get more biological diversity. You literally just sample it from a generative model. And so I think this is another thing that we're trying to kind of follow up on, but like the value of a generative model as providing additional diversity beyond the diversity that we have even observed in nature. And then, yeah, what was the other component of the CRISPR systems that you're interested in?

Well, yeah, maybe a better formulation of that question is, you know, we can now generate potentially unlimited CRISPR-like sequences. Is there anything other than actually synthesizing them and trying them that you could use to figure out which ones are good? And how practical is it to test like 1,000 or 10,000 or a million new CRISPRs?

Yeah, this is still fundamentally limited by like a lot of different areas of biology. So the throughput with which we can test these CRISPR systems is low. We can test maybe 10 to 15 or something in a reasonable amount of time and money. Where I see the models as being Really useful is that if you just take something from nature or someone tries to rationally design a protein, the success rates in the lab will be quite low. If you just told a human expert to mutate CRISPR so that— or mutate 400 amino acids in CRISPR to bust this patent or something, the likelihood that that would be successful is actually astronomically low, I would say. But maybe with EVO, you could get it as high as 10% or something, or 20% of the designs that EVO or a generative model generates actually do succeed in the lab. And the goal is to get this— it's called low-end in BioML, like this kind of low-end design— actually achieving this versus having to rely on brute force, high-throughput methods for screening designs, which are costly and very expensive.

Yeah.

get the models to help out with kind of low-end validation. So you just test 10 things and 1 or 2 of them works. I think that's already a big win in biology. But yeah, we're fundamentally limited by the biological assays to validate them. And it's very important. Like, I definitely agree. The ultimate test, because the model will hallucinate or just make mistakes, is to just test the thing in the lab.

So maybe the last thing on EVO, and then we'll move on to the 3rd paper. I know we have a time budget. is just on the big picture safety side. I noted in the paper that you guys say that you've excluded viruses that infect eukaryotes entirely from the dataset, which is a pretty significant exclusion. And that seems like a very prudent step for such early technology. So I applaud such a thoughtful approach. Big picture, it seems like This, like everything else, is going to get scaled up. It's going to get better as it gets scaled up. And then right now we already have a lot of kind of debate and concern around, can ChatGPT help me make a bioweapon? But it seems like the real sort of bioweapon generator, unfortunately, is probably the same thing as the medicine generator, because these systems are understanding things at such a fundamental level that it's It doesn't seem separable. Maybe you have a different perspective on that, but how do you think we sort of develop these technologies safely in the long term, knowing that one pandemic can be a big, big problem?

No, definitely. Yeah, we definitely care a lot about safety, like biosafety and biosecurity. We have people in the lab testing these things out. I don't want anyone in my lab to get sick. And definitely want to make sure that these technologies are used for the benefit of humanity rather than to make a weapon or something like that. So yeah, there's a number of things that we've been trying to do. One is in terms of data exclusion, like actually excluding sequences that are potentially harmful, like of these eukaryotic viruses. I think that was one step. I think just in future versions of the model, actually potentially developing a set of evals of common viruses or maybe even toxins and actually making sure that the model maybe has low perplexity on these sequences. Or if it does understand these sequences well, deliberately having the model in some kind of post-training stage, like actually not be able to reason out these sequences very well, the ones that are kind of harmful. Yeah. And in terms of the larger safety and security questions, I think probably the scientific community and policymakers and just members of the public should have established some kind of standards that make sense and can continue to enable progress in this area. But again, also progress that is responsible. I think I fundamentally take a very optimistic view of this technology because it's much harder to create or to ameliorate bad biology than it is to destroy. Like, unfortunately, if a motivated state-level actor wanted to create a bioweapon, they could do it quite well without generative AI already, actually. And in fact, I would not use a generative model to design a bioweapon right now. But I think what the technology does let us do that we just can't do well right now at all, because it's much more complicated to, it's much harder to create than to destroy, is to actually develop cures. Or vaccines for these diseases that nature has already been producing and releasing into the human population. It's much, much harder to design biology in the ways that we want and kind of return things to homeostasis or reduce the entropy in the world. And I think that's where I think these technologies will be really useful. And that's where the big capability jump, I think, is. with these models is actually in enabling us to kind of correct the natural tendency of biological systems to deteriorate and get worse. Like, it's much, much harder to kind of reverse that course. And I think that's where these generative models will make the most impact. So this is why, like, I'm optimistic about the technology. I think we should continue to develop it safely with the right guardrails in place.

And if we do that, then we can actually have a tremendous benefit on human So if I summarise that in terms of offence-defence, as it's sometimes put, it sounds like your point of view is the natural state of affairs favours offence. And maybe the reason we don't see offence is actually just lack of people out there who want to do bioweapon offence.

Yeah.

And this technology, in addition to all the obvious good things that it could bring about actually favors defense.

Yeah. Like, it's—

On a relative basis.

It's, if you want to make some biological system go off the rails and deteriorate or become diseased or, you know, cancerous or something, it's actually quite easy to do this already. Or you just, you know, you have an organism and you just like hit them with a bus or something and the thing dies, right? Like, it's very easy for these biological systems to deteriorate. It's much harder to exactly play like defense or to correct things and return things to like this kind of very low entropy state where the organism is nice and in some kind of homeostatic balance. And so, yeah, I do recognize that these technologies elevate our ability to do biology broadly for good and for dual use purposes potentially. I mean, this technology does elevate our level to do biology broadly. But the way I see it is our current level of technology now, like we've already cleared the bar for a long time actually, in our ability to make biology worse or to harm people or something. Where our bar is right now in terms of biological capability is like far below the capability needed to help people. So we've kind of already cleared the bar for hurting people for a long time, like, I don't know, nuclear weapons or something. Like, you just subject someone to a bunch of radioactivity and that also causes cancer, right? So our ability to make weapons, biological or otherwise, has already cleared the bar for technological capability. Where the capability level is not yet high enough is to actually fix things. That's why I'm optimistic about the applications of this technology.

Yeah, that's a great perspective. All right, we'll have to go through the 3rd paper somewhat quickly. This one is Full title, Unsupervised Evolution of Protein and Antibody Complexes with a Structure-Informed Language Model. So to kind of complete the trilogy here, you've got the first one is all about exploring the ML architecture space. The second one is about trying to get through just raw data of DNA sequences, start to get a handle on these higher-order concepts, and at least start to make progress on this. What causes what? What interacts with what? And then this third one is really about saying, okay, we've got a particular target, let's really lock in on that target. So you take a model called ESM-IF1, and there are a couple of things about this that I didn't quite understand, but this is a model that combines both sequence and structure data into its understanding of the world. And I was very interested to see that it was only trained on single sequence, and yet you get it to generalise to complexes where multiple different things interact. What I didn't understand is, did that just kind of happen? Is that like emergent good luck, or was that actually like hard to do?

No, that was totally Maybe somewhat unexpected, actually. It was a— I guess you can call it like an emergent property of the model. So yeah, the way the model works is actually perfect that there's like an ML audience. Like, it's an encoder-decoder architecture. So, but kind of like machine translation, where the encoder is a protein structural encoder, actually, and there's a protein sequence decoder. So it's almost like we're translating structure, like the sequence of structural coordinates Into a sequence of actual amino acid identities. So the, the, the, the, so for each token position, we, we actually give the model the 3-dimensional coordinates of the backbone atoms of, of the protein, and then ask the model to then predict our kind of next token prediction conditioned on the entire set of backbone coordinates. What the actual amino acid identities are that fill in that backbone or that are compatible with that backbone or that That will fold into that backbone if the sequence is actually produced. And yeah, the model is only trained on single chains. It's not even trained on multi-domain proteins. It's trained on very simple proteins. So actually the training dataset is like these small globular domains. So like a single chain of a protein can actually have smaller kind of units that are globular or that form kind of domains. More like nice compact units called domains. So the model is not even really trained on large multi-domain proteins. It's mostly trained on these domains of proteins, but it still generalizes to complexes. And the way we do that is actually very simple. We just treat a multi-chain complex as just a huge protein, potentially where the 3D coordinates have large discontinuities Or like large differences in coordinate space from each other. But we just feed that whole complex into the model and it seems to understand which mutations either stabilize the entire complex or destabilize the complex. And when you stabilize a complex, typically that involves improving the binding affinity between the constituent members of that complex. And when you destabilize a complex, you decrease the binding affinity between the 2 or more constitutive components of the complex. And this is very useful for, say, antibodies. So antibodies have 2 chains. So antibody has a heavy chain and a light chain. And then they often bind an antigen, which is one or more chains. So that minimum 3 chains involved in an antibody-antigen binding interaction. So we just give the model the whole set of 3D coordinates of the antibody bound to the antigen and say, you know, which mutations on the antibody side will stabilize that complex? And it turns out that a lot of those mutations helped improve the affinity of the antibody for the antigen, which is really useful, especially, for example, in clinical applications. You have some antibody that kind of binds weakly that you've isolated from a patient or from some initial library, and you just want to make it 10x, 100x better or something. So this is the kind of area of the molecular development lifecycle that This model fits into.

Yeah, it's another striking example, even I think more so than I realised, of a model generalising in a pretty significant way outside of the domain that it was trained on.

Yeah.

So let me just make sure I can articulate what is happening in very literal terms. Tell me if I get anything wrong. You start with a structure of an antibody bound to the thing that it's Evolved to target.

Yeah.

And you feed into the model the 3-dimensional coordinates of the backbone atoms of this protein complex. The model has only ever seen single-chain proteins that are folding up on themselves. You're feeding in 3 different chains that are interacting with each other. It doesn't seem to care. Its job is to predict based on those coordinates, which amino acid is at each position, and it spits out a prediction. And then you compare those predictions to the actual ones that you had, and you look for differences and you say, well, geez, if it thinks that it was this one when it was really this one, well, maybe the one that it thinks would actually be better. And then you can go test that. And then if I understand the sort of loop correctly, it's like, Take the ones that have kind of the most likely— where the prediction from the model is most different from what is the actual, go synthesize those, test and see which ones actually do bind better, then come back to the model and start with that new one and run it again. And that way you can do 2 rounds of evolution. And in the end you get like significant, depending on exactly which complex, of course it will vary, but order of magnitude or more stronger binding. I actually don't know what that means. So when it's like 25x improvement, does that mean like I'd have to pull 25 times as hard to pull them apart? Or like they stay together like 25 times more of the time? This is how naive I am, but tell me if I got the setup and the loop right. And then like, what does that actually mean when it's 25 times better?

No, the setup is Yeah, it's almost there. Yeah. So we give the model just the setup. We assume that we have the 3-dimensional coordinates of the antibody bound to its target, or I'll say antigen, but that's the jargon, like the thing that the antibody is bound to. And this is determined with an experimental method, typically cryo-electron microscopy or X-ray crystallography. We don't have to go into the details, but they give you a pretty good resolution picture of where the backbone coordinates are. And then we also know the sequence of the antibody and the sequence of the antigen. So actually in our decoder, that decoder part of the model, you actually kind of, for the parts that you don't want to change, you just fix those or just score the actual amino acid identities that are present in the original or in evolution, here's the wild type starting sequence. And then what we do is We then just make single amino acid changes at each position in the antibody and then see how does the likelihood of that entire, you know, just the log-probs of that model change when we make this mutation compared to the wild type or the original sequence. So the idea is, you know, if the likelihood of some mutation gets really high, then maybe that sequence is more compatible with that set of backbone Whereas if the model at position 10 thinks that an alanine has 0% likelihood of appearing in this position, then maybe if you put that alanine there, it's going to completely disrupt the structure, for example. So that's basically how— and then we just score all possible single amino acid changes to the antibody. So it's a sequence length times number of amino acids, number of changes. And then we just rank those by how much they affect the likelihood under the model. So at the very top of that list are things that, you know, potentially are going to improve the likelihood of that sequence under the model.

Yeah.

And then so the loop is in the first round of evolution, we just picked the top, I think like 20 or 30 mutations by the likelihood ratios and then tested them individually. And then the second step of evolution, we just do something Kind of silly, or, oh, I mean, not silly, but very simple, is just make various sets of combinations of these mutations. And we can also score them by the model itself. So we just make all possible combinations of the 5 or 6 mutations that actually improved affinity or improved binding in the first round. So the second step is a little bit more manual. We're not retraining models on that data and then updating the actual weights of the model or anything. We're just kind of combining, though we can use the pre-trained model again to score different combinations. And yeah, by improving binding by some 100-fold or something, or 100x, this is fundamentally— so the actual measurement that we use for this is actually a concentration-related measurement.

Okay.

And you can think of it as like the concentration at which, in terms of like the chemical reaction in molecules A and B. It's like the concentration at which there's an equal amount of A dissociated with B and A bound to B. So, it's kind of like an equilibrium, but fundamentally it's a concentration. So, at this concentration, there's an equal amount of molecule A dissociated with B and A bound to B or whatever.

So, 25 times better means I only need 1/25th as much to achieve that. Exactly.

So, it's a concentrate. Yeah, it's a concentration-based readout. So, and this thing could potentially translate downstream into like dosing, for example. So, if your molecule you know, works as well at 2x the concentration. Maybe you actually, and you have that your activity is good enough, maybe you just dose at, you know, half of the actual molecule. Yeah.

Yeah.

Cool.

Okay.

We're at time. If you have time for one more, I'll give you one more. Otherwise, I can thank you and send you on your way right now.

I'll take one more question. Yeah.

Okay.

Well, you pick which one you want to do. My last 2, you can do one or both. One was, When you're doing this sort of superfitting, do you have to worry about overfitting? Because obviously the target is subject to evolve and change too, right? So do you have to worry that now that I fit this thing so tightly, if it changes a little bit, now that might throw us off again? How do you think about that? Is there some sort of many— do you do like a Monte Carlo tree search type of approach to try to game that out? That's one question. Other question is, Interested in your comments on the Ark Institute. It's coming up all over the place. There's like next generation CRISPR successor recently announced out of there, and it seems to be quite a special phenomenon. So you can take one or both of those.

Yeah, I'll just answer both very briefly. Yeah, in terms of like, you hit the nail on the head of what I am pretty interested in moving forward with this kind of stuff is that in therapeutic design, we're often assuming the But the target is fixed, right? And it's not changing over time. But actually in nature, definitely in viral infectious disease or in bacteria or in cancer, the target is a moving target and it's actually constantly evolving and will constantly evolve around our therapies. So you spend $2 billion to develop the drug and then you administer it and then 6 months later you get like drug resistance. But I think there's a lot of room here for interesting algorithms. You brought up some kind of tree search or maybe some kind of RL policy or something, or more basic Going back to the original set of questions, for the second point, you know, it seems like we made a lot of progress in molecule design, but I think this is one area where the algorithms could be pretty interesting. Actually, we made a lot of progress if you're designing against a static target. What happens if it's a moving target? And then ARC is great. I'm at ARC right now. I'm actually on site. It's a really great environment. They've been really supportive for my research and just I'm just starting off my research career. So I've been technically a professor only for 8 months now. And I don't think we could have moved this fast without like the support of ARC and the immense like talent density here, I think. And yeah, I mean, it's great. I think there's definitely a lot of room for innovation in models for funding science and supporting and just doing science. And I think the ARC model seems to be— Doing well so far, but I know we'll see in time. But so far, I think it seems to be a great place, and I'm especially struck by the collegiality among the people here. Like all the investigators here are very friendly, and we all collaborate, and I have projects ongoing with all the other investigators at Arcor in some way. I'm involved in a project with them, and so it seems like a great place for now, and hopefully in the far future. As well. Like, there's no reason for me to think that it won't stay this way. But yeah, I think it's been good and our park has been great.

Cool.

You guys cure cancer, you'll deserve all the fame and fortune that that will bring. And leading indicators are definitely there that special work is coming out of that place. No doubt.

Yeah, it's, it's, I've been very fortunate to be in this environment. So yeah, I guess I am technically sponsored and that they fund some of my science, but that ad was not sponsored.

Well, it's been a remarkable series of papers that you've been involved with, and I really appreciate the time to walk through them with us. I am sure there will be more great stuff to come from you, so we will be keeping an eye out for that. But for now, I'll say, Brian He, thank you for being part of the Cognitive Revolution. Cool.

Thanks, Nathan.

It is both energizing and enlightening to hear why people listen and learn what they value about the show.

So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.