#202 Raphael Townshend: How AI and RNA Tech is Transforming Drug Discovery (Inside Atomic AI)

29:48And so in some ways, we have these sort of top tier AI folks working, but then we're also building the right data that's purpose built for these kinds of models. And this is really sort of a broader trend that I'm quite excited about, which is kind of this integrated lab sort of computation kind of approach where you have this iterative cycle where you're generating data that then makes the AI better that can then feed into more data that's more targeted to further improve the AI to get that kind of virtuous cycle of improvement. Yeah. So can we take the example of an undruggable protein and you understand the mRNA that generates, I guess, is that the right word, that protein or builds that protein or has the instruction set to create that protein?

30:46and you want to interrupt that mRNA molecule. The first step then is to understand the shape of the mRNA molecule. Yeah, so at that point, what do you do? Yeah, no, it's actually quite interesting because in theory, you don't need to know the shape to start trying to go after it. You could just take your RNA and start throwing molecules at it and see what happens. And in fact, that's what, you know, the first generation of companies in this space did, like about a decade ago at this point. And the issue there is that the molecules were generally not potent, first of all. They didn't do what they needed to do, right?

31:29It would stick to the RNA, but it wouldn't, like, destroy it or decrease the amount of protein, et cetera. And on top of that, they were oftentimes not selective. So they bound to that RNA, but they also bound to many other RNAs at the same time. So there's issues in potency and selectivity. And that's really where the shape starts coming into play. Because you want to find kind of the unique 3D shapes, the unique pockets where you could get a molecule to stick there, and then you could optimize it to bind to that location, not others. And on top of that, you know, if there's a well-defined shape somewhere, nature doesn't waste effort generally.

32:05And so those things are generally the pieces that are actually functional and doing something interesting versus like some generic piece of RNA might not actually be doing kind of any interesting function. And so shape X then becomes critical to get over these barriers of function and cell activity. And so to give you an example of what this does then is you could find a shape in the RNA molecule that's responsible for keeping it stable. I mentioned stability before. And you can sort of destabilize that structure with a small molecule. It's kind of like it binds in there and it makes it less structured, basically.

32:41And then that lets enzymes, right, other proteins that are responsible for chopping up this RNA molecule have an easier time to doing so. So you're sort of destabilizing it, making it more prone to getting chopped up, and that decreases the amount of it, decreases the amount of protein, you know, cures cancer. But how do you discover the shape? I mean, that's essentially what you're doing, right? Exactly. So, and this actually gets at the reason why folks are so excited about things like alpha fold in general, because the traditional way that you find these shapes is through these very expensive, very slow experimental techniques.

33:23They have names like X-ray crystallography or cryo electron microscopy, etc. And I mean, we could get into the technical details of how they work, but really the key to remember is that these things can take months or years to solve a single structure. Okay. I have a good friend of mine. He spent his entire PhD solving a single protein structure using these techniques. Okay. And these machines, like the cheap machines in some ways, cost millions of dollars. um and so if you can take this process that takes months or years to get a single shape and then use these ai approaches to bring it down to minutes or seconds instead that's a big deal like people really care about that basically um and so what we're doing right is we can you know instead of relying on these expensive techniques and trying to run those over time which you know over the past couple decades have solved maybe a thousand rna structures total depending on how count.

34:19That's actually generous. Maybe it's a couple hundred. You can instead take these AI approaches and just map out everything at once and find these interesting structures. You're finding those sort of the pieces that are folding into nice pockets that are targetable in the first place. And so one way that I talk about what we're doing is we're essentially taking the space of all the RNAs in your body and identifying which parts of it are the most sort of like structured and targetable through drugs. Well, okay. And so how do you do that? I mean, you have, is there a library of mRNA shapes already?

35:00Or are you? You're asking how you train those kinds of models in the first place or something. Yeah. So there's an existing library of shapes, right? That have been solved through these expensive experimental techniques, as I was talking about before, right? Like maybe you have a couple hundred of those, right? So that's the starting point. That's your sort of gold standard. Now, as I was just saying, these AI approaches, right, are very data hungry in general, right? And you can get some initial bang for your buck through being clever in the algorithm design. That was the initial science paper.

35:36But eventually, you sort of run into the bitter lesson of like, You just need more data and more compute to really get over some of these things. And so this is actually where sort of RNA itself ends up being quite an interesting molecule because you can design experiments that are very high throughput, that give you lots of measurements in parallel for RNA specifically. And that's because you can connect it to DNA sequencing. And the cost of DNA sequencing has fallen off a cliff over the last couple of decades. And so we can design these experiments that can measure tens of millions of RNAs in a single shot.

36:12And so let me explain actually how one of these techniques works just to paint the picture for a second. So you have your RNA molecule, okay, and you expose it to a chemical, okay? And that chemical will go and like nick the RNA at different points. It'll kind of damage it basically, okay? And so then you can convert that pretty easily back to DNA and then you can run it through your DNA sequencer. And the parts that got damaged, basically, will show up as mutations in the DNA sequencer. Okay. So it's kind of like, it'll be like errors in the DNA sequencer. And so now you can sort of very easily pick up where these chemicals damage the RNA molecule.

36:48And the key is that the places where it gets damaged are very linked to the shape of that molecule. Like one very simple way of thinking about it is that the outer parts of the RNA are going to get more damage than the inner parts. Okay. And so now you've got these measurements that are telling you something about the structure of these RNA molecules. And because it's DNA sequencing link, you could just run this in parallel on a huge number. And on top of that, it's actually really easy to make at least shorter RNA molecules at a very large scale. You can sort of, this process known as oligosynthesis lets you just do it synthetically and create like millions of them at once.

37:25Versus for proteins, you can't actually sequence those. You can't run those through DNA sequencers very easily. And on top of that, making them, it's harder. You kind of have to have the cells produce the proteins for you. And so in some ways, the fact that we're operating at the RNA level makes this job easier because you can make and measure the RNA molecules at a much higher throughput than you ever could for proteins. And can you then model the RNA shape in visualization the way that you can with AlphaFold? Exactly. Yeah. So you end up with a shape of an RNA molecule. And I could even show you some or something like that.

38:08But it's basically like how I think about it is you just like an RNA just native atoms at the end of the day, like a protein or anything else. And it's just the atoms in 3D space. It's like, okay, you've got a carbon over here, you've got a nitrogen over here, and they're bonded together, right? And so you can look at those and be like, okay, that's, you know, you can sort of see how it's structured, and you can really even start simulating it. And you can just run the laws of physics on that, right? And that's another process that's super interesting and can tell you how it flops around over time.

38:41And then to create the small molecule to bind to the RNA, is that also then once you understand the shape of the RNA molecule, do you search through some search space for an existing molecule that has a corresponding shape? or do you then synthesize a unique molecule? Right. Yeah. I mean, there's a number of different ones you could do it. One of the common ways is really known as this process known as docking, basically, where it's like you've got the shape, right? And you're just trying different molecules and saying, like, do they fit? And does it interact physically well with the other molecule, right?

39:31And so you could search through these very large spaces of possible molecules, including things that have never been synthesized before, and then say, this one looks good. Let's go and make this one, and then actually test it in the lab. So the idea is you're kind of searching through this massive space and then narrowing it down. And that's one common way. There's a few different ways you could use these things, but fundamentally, you've got the shape, you understand what you're trying to go after, and then you're trying to find molecules that interact with that shape. And which do you do? And which does Atomic do?

40:02Yeah, so we do a lot of this docking type of approach, as I mentioned, of fitting the molecules in there. And then we combine that with more traditional kind of screening methods as well, which let you sort of, you can sort of, once you've identified where the shape is, you could just isolate that shape. And then you can throw a bunch of molecules in a lab setting at it as well. And so generally, we apply both those techniques together, like the joint computational experimental piece. And you can even combine those together, right? Because once you, if you've run it in the lab, you can then feed that back into your AI and do a better job of docking, for example, the next time around.

40:38Yeah. And you were talking about transformer models. So this isn't search and reinforcement learning. You're generating molecule shapes. exactly and and then if you find one that computationally seems to fit then you synthesize it and and test it in yeah yeah you find that you've you get some rna's you you generate their shapes you screen small molecules against it computationally if something looks promising you go and synthesize that and you test that first in cells and then in animals and you keep pushing towards the clinic on that front. Yeah. And then eventually into human trials, I presume.

41:32So where are you in that whole process? Where is Atomic today? It's actually kind of an exciting time for Atomic because we're just starting to test in animals for the first time here. So, you know, I would say it's still what I've called the preclinical place. We're definitely early stage in many ways. But this is really the first time that we're going beyond cells. We've seen our technology work well within cells. And now we're trying to get the next sort of layer up, the next higher level organism in some ways. And really putting sort of a lot of the platform to the test. And we're anticipating getting a lot of that initial data in the not too distant future.

42:14So it's really, I don't know. I'm personally, it's a cool moment because, you know, I've been working in the space for like 10 plus years at this point. And it's like, okay, we're actually like, it's starting to be the dream starting to become a reality in some ways. Yeah. What, what, what happened with, um, I haven't looked at AlphaFold 3 yet. I just saw the announcements. Uh, what, what is the, what's different with AlphaFold3 and are you adopting whatever changes they made in AlphaFold? I've forgotten. Did you call it AlphaFold RNA? Yeah. Yeah, no, it's, I mean, our model actually, our core model is known as Atom1.

43:01It's an RNA foundation model is, I guess, what we call that. I should plug the name a little bit, I suppose. But, you know, I think we're pretty excited about AlphaFold3 overall. You know, I think the space as a whole has been moving really quickly. Right. So there's always these new advances coming from different groups. And you always, you know, make sure you read through the papers and understand and integrate the pieces that are useful from these different advances. I think the key that's happening with AlphaFold3, right, is that they've expanded like their modeling from just proteins to a much broader space of molecules.

43:34Right. So that includes RNA as well, to be clear. And they've seen some pretty good success at at least some of these new molecules in doing this kind of modeling. For example, they've seen pretty good success at modeling protein-small molecule interactions, like a little bit more like what in silico medicines might do, right, of modeling protein-small molecule interactions. They've done a pretty good job at DNA or things like that. However, RNA actually is one of the areas that's still room for improvement, at least based on their studies. because they still don't have state-of-the-art as compared to more traditional methods there.

44:10And fundamentally, that comes down to the fact that there isn't that much RNA data out there that's public, right? And this is kind of one of the key bets of Atomic is like you can do very well in certain areas where there's a lot of data, but in others, you really need to invest carefully in collecting the right kind of data as well. And so I think there's a lot of very useful components that have come out of the AlphaFold3 kind of approaches. And we're definitely, you know, it's a fun paper to read overall. And I know a lot of the team members there quite well, and I'm pretty excited for them.

44:44And it's really like taking those pieces, combining it with the data that we have already that we spent the last three years collecting in-house to really try and create these continued breakthroughs in the RNA space. Yeah. How many RNA molecules have you modeled successfully? I mean, to the point that you can synthesize molecules that will fit or bind with them, whether or not there's therapeutic interaction. How many have you done so far? Yeah. So it's a super interesting question. So I would say there's different levels of validation you can do. It's like, okay, you can make the molecules and then you could test them in certain ways.

45:32And one answer to that is tens of millions or hundreds of millions for the highest level validation, or we've made these things and we've tested them, found their structures, etc. But then it's like the number of things that we're testing in animals, which is several steps later down the process. We're just getting our first one there. So you could think of it as a funnel that's starting from this huge number and then where each step validation is getting smaller and smaller, but it's the tip of the sphere there. Maybe one stat that I like to give is there's a level of accuracy you need to do this rational design approach.

46:12It's like you want your structure to be this close to correct for you to be able to model the interactions with a small molecule that could buy in there. If your pocket is completely wrong, it's not going to really help. And what we've seen over the last couple of years is that the median structure, on average, the structures that we're making are sufficiently accurate now to be able to do that rational design kind of approach. I'm not saying that they all can do it. But it's like the original science paper that we put out. Maybe it was like 5 % or 10%. Don't quote me on that. I don't remember the exact number, but it was a small number, basically, a relatively small amount there.

46:53But since then, we've brought in over 50 % of those structures are now sufficiently accurate to do this kind of approach on it, which personally is kind of the holy grail in some ways, as far as I'm concerned. It's like, wow, okay, we can actually use this reliably. Yeah. Yeah. And once you get the process down, is it a matter of just, I mean, running this iterative loop until you reach a level of accuracy that it's worth going into it. Yeah, no, it's, yeah. No, yeah, exactly. It's like you need the right level of accuracy. And I don't want to overly simplify either, right? It's like you can definitely turn the crank and eventually you'll get there.

47:40The one thing that's interesting and, you know, difficult in biotech is that biology is complicated. And there's a million and one things that, you know, could go wrong, right, in various ways. Like you get a molecule that's very potent and selective, for example, but if it doesn't circulate well through your body, then it's not going to do much anyways. And so there's a lot of pieces that you need to put together. And it's like, we're not replacing the entire drug discovery and development process wholesale here, but we're really honing in on some of the key aspects, some of the critical bottlenecks that have hit the field and made those better.

48:14Like we're not replacing animal testing, as an example, right? Or, you know, we're not replacing the clinical trials themselves, but we are getting to faster, better molecules to run through those things. Yeah. And your focus right now is on narrowing to molecules that you can test, or is it on the sort of the more general, as you were saying, the dream is to eventually model all RNA molecules? Or are they happening in tandem? Right. Yeah. I mean, it's really in tandem in some ways, right? It's kind of like how I think about it is you've got the long-term bets, right, and then the mid-term kind of things and the short-term kind of pieces that we're doing.

49:02And I think especially for a science-heavy company like Atomic, you need that sort of balance, right? Because you want to be looking at that near term of like, let's actually show this thing can deliver on the promise in some cases, but then also sort of enable that much broader sort of space at the same time. And so we continue to build what I describe as RNA foundation models, collecting these very large data sets to build accurate models of RNA structure, of RNA function, et cetera, enable RNA design. Like that's an active area of research here at Atomic. But on the other hand, we're also advancing our first programs into testing in animals and narrowing the search space, as you say, to find those molecules and testing them.

49:44And, you know, I think this is kind of a little bit my own philosophy as well, which is you want to be applying the technology that you have or like putting it to the test as much as possible. Right. Like we're not, you know, I'm a big believer in building the thing that is useful, that'll actually move the needle as opposed to, you know, designing it sort of in isolation and then figuring out how to apply it. Because I think that by applying it and looking at where it's useful versus not, then you can guide further efforts in that direction and really build the sort of the foundation models that are useful and that are really going to make a difference.

50:23Yeah. But the activity, I mean, you've got this animal trial. If it's successful, do you then move into human trials? And is the, I mean, I understand the long-term goal, but is the company at this point focused on generating data that then can train better models? Or is it, I guess I asked this already, or is it on coming up with therapeutics? Yeah, I mean, the boring answer is really both, frankly, is what I'm getting at in some ways. And these two things are very linked in some ways, actually. I think you're getting a very important point, though, because these first trials that we're doing, where we're testing in animals, for example, that's going to generate a small amount of data, right?

51:15And you can only feed that back in. But oftentimes, the kind of data that's really critically useful for training these big AI models looks fairly different than the data you get out of any given drug discovery program. And so, I mean, maybe concretely the way I could paint this for you is that we have a team that's dedicated to generating data for the AI models specifically. And then we have another team that's dedicated to pushing forward the drug discovery programs. And there's a lot of cross-interaction, cross-pollination between those two, but there are specific folks with very specific mandates at Atomic along both those lines of what you just described.

51:52Yeah. And remind me, when did you form Atomic? Yeah. So it's been three years that we've been going now. We're like 25 people today, half sort of AI scientists, software engineers, et cetera, half sort of RNA biologists, medicinal chemists kind of side of things. And it's really, I've been saying this both thing a lot here. And it's like, fundamentally, it's kind of like we're trying to build this interdisciplinary organization that has that expertise across these different spans. And we're all sort of co-locating the Bay Area to enable that kind of interplay and interchange of ideas. So I fundamentally believe that a lot of the key innovation for a place like Atomic happens in that white space between established fields in some ways, right?

52:39It's like, you want to build that new field there at the intersection. Yeah. I was just in DC talking to the director of the National Science Foundation. How much are you depending on government grants? Because this sounds like something the NSF or other government organizations would want to fund. How much are you depending on venture capital? Yeah. So at this stage, we're mostly venture capital based. However, you're completely right. There's a major sort of government angle to this whole thing. In fact, actually, it's funny that you asked the question on government. I literally met Secretary of State Antony Blinken on Monday talking about AI applications for biotech specifically and making this argument that some of these sort of large data sets you need to collect to build these foundation models, right, are like really key, right?

53:35It's like to build the best foundation models, you need the best data sets that require sustained investment, long-term investment, the kind that the US government is uniquely suited to provide. And so you're actually hitting on a very interesting point there that I think that is very much an area of excitement for atomic and sort of this AI for biotech's field in general. Yeah. And I didn't realize that ARPA, is that how you pronounce it? Yeah, IARPA. Now there's IARPA H. Right, ARPA H, right, yes. It was focused specifically on healthcare applications. And I also didn't realize until this conference that there is a National Security Commission on biotech now.

54:26Are you involved with them at all? I haven't been involved to this date with some of those discussions. I have been involved with others. I would say that the way that I would describe it is there's this realization across the government in some ways that biotech is a major area of innovation that needs focused sort of investment. And I think that the U.S. wants to continue being sort of a leader on the global stage there specifically. And so it's thinking through carefully how to back these kind of things and how to potentially increase that investment there. So I think there's a lot of interesting conversations.

55:04I mean, I mentioned the one I had on Monday on this front as an example. And especially in light of these new AI applications, right, you've had these executive orders, right, get signed, right, from the Biden White House, right, sort of really increasing the sort of focus on these kinds of areas overall. Yeah. And do you have any trouble with funding or is there plenty of money for initiatives like yours? I would say AI continues to be a very strong space for investment as a whole. I think the biotech market specifically has been going through a bit of a rough patch since the pandemic. However, AI for biotech is the bright spot across that landscape would be one way that I think about it.

55:53Because it's really showing like, hey, we've seen what ChatGPT can do. We've seen what AlphaFold can do. There's a lot of excitement in some ways. And it's almost like a reverse hype kind of thing is one way of thinking about it. It's like the people that are closest to it are oftentimes the ones that are the most excited about the whole thing, which is kind of cool to see, especially on the AlphaFold side of things. Maybe chat GPT now everyone knows about that and is excited about that already. I don't want to say that that hasn't gotten into the public consciousness at this point. Yeah. Yeah.

56:26Well, I was to that point, and he's on the commission. I can't remember. He's now the founder of Ginkgo Bioworks. Very nice. Yeah. He was saying that he thinks within two or three years and maybe it'll be atomic AI, there's going to be a chat gbt moment for for uh biotech that that's yeah um i would agree with that i think that we're going through there's going to be some really exciting developments in the next few years yeah uh is there anything i haven't asked that i should ask we're up to an hour yeah i mean i think this is really good i mean i i covered most of the points that i wanted to hit here.

57:10You know, it's like trying to integrate the wet with the dry lab together to make that cycle, you know, this RNA breakthroughs, why we should care about RNA specifically is key. Atomic is really trying to bring RNA and AI together to usher in that new generation. No, I think we covered it actually. Great set of questions overall. Yeah. One question. What kind of compute demands do you have? And is there plenty of compute for what you need? Again, I had this conversation with a guy named Brian Spears at the Lawrence-Liverborn National Laboratory, who's doing work on RNA shape discovery. Very cool.

58:02But they've got the most powerful computers in the world at their fingertips. How are you handling compute needs? Yeah, no, it's I mean, everyone, I think we always need more computers, really, the honest answer. We've got, we've bought a bunch of GPUs that we have on premise, basically. So we always have access to, and that lets us sort of do a baseline of like work, basically, over there. And these are sort of like H100s, basically, which are like the top of the line GPUs that you really need for this kind of work. And then on top of that, we burst up the cloud. When we're doing a big training job, then we go to AWS or GCP or what have you and train there.

58:43Now, the issue is always getting quota on these things, getting enough allocation on the cloud providers to run these kinds of things. Because in theory, they have a lot, but you can't always get access to it if there's a lot of folks training at the same time. So we're always hunting for more compute is maybe the simple way of putting it. I used to work at the Department of Energy a little bit during my PhD using the Summit supercomputer, which I don't know, some ridiculous amount of H100s. And it'd be very nice to just have that at our fingertips at some level, like 27 ,000 H100s or whatever it was.