🔬The BioAI Phase Shift - Matthew McPartlon & Neil Patil, Chai Discovery

19:40Neil Patil:We're like, all right, this is our forcing function. We have a clear goal to work towards. It's very direct. Let's get this thing going and see how fast we can do it. You guys were at this time sitting in the OpenAI offices? We were sitting in the OpenAI offices, yeah, in the mission. Right. So what's the backstory on that? It's really interesting. Two of our other co-founders, Josh and Jack, had a relationship with some of the OpenAI people. Actually, OpenAI co-led our seed round. So we were kind of thinking, all right, should we get an office while we're only five people? And it turned out that office was mostly vacant.

20:10Neil Patil:So we got to sit in the OpenAI offices for a while. Chai One built it, open source, learned about infrastructure. Yeah. So then after that, we really set the sights down on protein design.

20:23Matthew McPartlon:And worth pointing out, Chai 1 was a structure prediction model, right? So you have the sequence, what is the structure that it folds to? And then that was Chai 2.

20:32Neil Patil:Yeah, Chai 1's finished. One other crazy story there. Let's see if we can actually share this. But this is a hilarious one. So we were like, oh man, we really want to be the first to put this out. And we were like, okay, we're one week out. We're like, the model's almost done training. We're like, should we build a web server? And then we're like, oh yeah, maybe not. And then like we ended up spinning up like this whole web server so like people can use it. Like rather than just like download the Git repo, it's kind of annoying, especially for biologists. And like we actually wanted people to use this.

21:01Neil Patil:So like let's spin up a web server. Let's get the technical report out, all this stuff. So we ended up like we were up for like 48 hours straight. It's like getting the paper over the line, getting like all the last things done on the web server. And then Josh was interviewing with like Bloomberg TV or something that morning. And we've been up for like 48 hours straight. So Josh runs into a room to do this interview on Bloomberg TV. And I think it was like 7 in the morning. Everyone's in the office. We didn't want to be seen or whatever. And the interviewer's like, oh, interesting company. Doesn't look like there are any employees here.

21:37Neil Patil:But yeah, it was a really fun time. I think the early startup days were just super fun. So yeah, after that, we kind of set our sights on design. And really what we were thinking is like, we kind of always had antibodies in mind. We thought of this as like the most tractable problem. The nice thing with proteins is you have this beautiful sequence representation. There's already a lot of research been done in like, how do you autoregressively generate sequences? How do you like the sequence generation problem is well studied. So we were thinking like, what's a nice like area to apply sequence generation to in the biospace?

22:10Neil Patil:And it's pretty natural to do like linear sequences of amino acids. So we start working on design. A unique thing about CHI is like, we're not like, we're designing antibodies and like we're an antibody company. Like we don't really like pigeonhole ourselves into like one therapeutic area. So we like tried to really tackle this problem very generally. So we were thinking like, can we design many proteins? Can we design antibodies? Can we scaffold regular complexes? So like really just take a holistic view on like how do you design proteins in general? And that eventually led to the CHI-2 model.

22:42Neil Patil:So that was our first flagship design model, and that's where the CHI-2 paper and our bold target discovery project came in. So we designed antibodies to 50 targets for that paper, got binders to about half of them with, I think, on average, around a 20 % hit rate for binding. And then afterwards, started working on CHI-3. So that's our latest series of model, but I'll break there. Before we talk about CHI3, can you tell us about, especially for listeners that may not be familiar with structure prediction models, what does the model look like? How does it work in general? Let's take a look at CHI1.

23:18Neil Patil:CHI1 has this like roughly a tokenizer, a transformer, something that looks like a language model, and then something that kind of looks like an image diffusion model. And they're all just like stitched together. The tokenizer is like not your kind of typical like words of X style tokenizer. This is like I have a bunch of atoms in a molecule, and now I want to like pull those into what I would call tokens for my like LLM-looking trunk. And then that conditions this like kind of big diffusion model, which will then emit the image, which is some 3D structure. So is it atoms or is it amino acids that are the input?

23:52Neil Patil:It's an interesting question as well. So we have like all these different input tracks. So like one thing about biology is the data is inherently multimodality in a sense. You have this token sequence representation. Each of these tokens has a set of atoms that dangles off. And then you also have some properties of the different atoms. An atom might have a different charge. It might have a different element type, so periodic table of atoms. And then these all get bunched together into tokens. Once tokenized, you can process this in very standard ways. But then ultimately, you have to get back to these 3D coordinates.

24:30Neil Patil:And it's like in order to predict the structure, this is just some 3D object. And that object goes through, or like to emit that object, you go through what looks like an image diffusion model where you kind of go back from tokens back to the atom representation. I see. So the tokens go in, the transformer establishes the relationship between the different tokens, and then the diffusion model turns that latent representation into a 3D structure. That's exactly right. Yeah. Okay, great. So that's CHI 2? That was CHI 1. Okay, CHI. So try one folding model. Yeah, it's like, and like all this bio stuff, it sounds like kind of scary, like atoms, tokens, amino acids.

25:14Neil Patil:Like at the end of the day, my background personally is like theoretical computer science. That's what I spent like all of my earlier years doing, transitioned to this like pretty late in my PhD. But I think like the background that you need is really similar to the background that you'd need for like any other field of machine learning. There are all these domain specific things that you learn about. but like one analogy or like anecdote I like to say is people think you can't work on like AI bio unless you're a biologist but it's kind of like you can't work on like video models unless you're like a director or something.

25:45Neil Patil:Like there are all these like super domain specific things like oh yeah to understand like lighting in a video, things like that but at the end of the day these are just machine learning problems and like they're all solved the same way. Okay, so then Chai 2, there's a jumping capability as well as an architectural change, right? Yeah, what we've disclosed about Chai 2 is like it is an all-atom diffusion model. So we're trying to predict, like, you know, atoms in 3D space still, but we're doing it in such a way that, like, the model actually has the ability to, like, design atoms, place them, decide which atoms actually are there.

26:17Neil Patil:So, like, one way to represent an amino acid, like a protein token, is by, like, which atoms are present. So in the CHI-2 case, we were just predicting, like, all right, show the model, let the model just kind of pick what atoms it wants to keep, and then map that back to what amino acids there are. What are you able to do with CHI-2 that you can't do with CHI-1? Is it just like better or are there new capabilities it brings?

26:41Matthew McPartlon:It's design, right? So CHI-1 lets you say, hey, I know the sequence of amino acids, right? That text string. And I know the structure. That you would get from like the genome. Right. Exactly. CHI-2 says, okay, I have a target structure, right? That I want to design a binder to. will then generate candidate molecules, candidate medicines that bind to that target. And so this is kind of a design model or design family of models. And I think that's where you really cross the threshold of usefulness, right? Like, I mean, CHI-1 alpha-fold, very useful because you can at least intuit and reason about the structure and see what you're looking at.

27:18Matthew McPartlon:But the ultimate goal here is to design medicines, right? And design new molecules. And I think CHI-2 really crossed the threshold of performance for doing that with antibodies a year ago.

27:27Neil Patil:one analogy here would be like kind of like back to like the image domain. So like Chai one would be like, you know, there is a cat in this image. Like, thanks, Chai one. And Chai two is like, I like, I'll show you a background, maybe like I'll prompt you with some, some like image information, like, hey, put a cat in a field and Chai two will actually just like give you back an image of a cat in a field. And you're like that, that's a good looking image or it's not, you might have some other model which kind of ranks the image. But fundamentally, it's the generative problem.

27:55Matthew McPartlon:So there's taking that analogy a step further, it's maybe more like you showed a background and then it generates, there is a cat and then it generates an image of the cat at the same time. And it makes sense that there is a cat in this field and also that the cat works in the image. So there's a, it is a, it's an interesting problem because you have to generate two things at the same time, both the sequence and the structure. Then you, if you, I don't know if you can, but could you talk a bit about like how that works? Like how do you do that? So you code, you co-design the sequence in a way that the structure also fits and makes sense.

28:27Neil Patil:One way to think about it is kind of like the classic way of doing this. Let's talk about both. And structure prediction, like, all right, I know the sequence. Like, I can from that roughly figure out the 3D shape. And then there's kind of like the inverse folding problem, which is like, given a 3D shape, give me back a sequence that would fold into this. And now you kind of like need to do both things at the same time. But I think like similar principles apply. Like, you can kind of have the model, like, think a little bit about what should this structure look like. Then you can have some other part of the model thinking about like, well, now what sequence would maybe support this?

28:57Neil Patil:And then like a nice thing with diffusion is like you can do this pretty slowly and pretty iteratively. So you can give the model a lot of time to think about, all right, if I change the structure like this, how should the sequence change? And you can kind of just play this back and forth and back and forth. And eventually it ends up kind of converging on something that's self-consistent. It's almost like an EM algorithm. Yeah, exactly.

29:17Matthew McPartlon:So you have this model now, Chaitu, which is able to predict or to sample a structure and a sequence which generates that structure. And just because you can generate a structure, like, doesn't necessarily mean it's necessarily accurate enough to do something. So do you have other scaffolding on top of that? Are there additional problems? Like, are you one-shotting these things or are you, you know, needing to generate thousands of them and then you have a ranking or scoring? Or, you know, how, like, Like just having a candidate is maybe, let's say, not enough. So what do you do once you sample a structure?

29:55Neil Patil:Traditionally, what's done in like when co-design and like protein structure design like started to become a thing, we're like kind of at a loss for metrics. It's like, how do you know that your protein, like you design some like sequencing structure? How do I know that this is legit or not? Like I can tell you it's like anything.

30:12Matthew McPartlon:It's like totally out of domain now, right? Exactly. That's my definition. Yeah.

30:15Neil Patil:Yeah, and like as a human, you can look at this thing and be like, I don't know, it checks out. Like even biologists are like, I have no idea if this thing actually folds. Like maybe some of it looks right. Even our biologists are surprised, by the way, with like some of our designs that like do end up working. What was done at the time is like we kind of came up with a bunch of metrics and like alpha fold, it really is what enabled this. So you'd take the sequence that you predicted. You'd run that through some like totally like distinct structured prediction method. So this is completely independent of your model.

30:44Neil Patil:And you say, if an independent model thinks that this sequence folds to a similar structure, then it has a higher likelihood of being correct than just whatever the prior likelihood would be. So you can take your sequence now and you can measure how consistent is this structure prediction method with the structure that you actually predicted for that sequence. You can now compare your design to an independent model structure prediction. And that became a really good way of gaining conviction that your design model was correct. and people kind of like game these benchmarks for a while and kept pushing, pushing, pushing.

31:15Neil Patil:It turns out like it's easy to get self-consistency, consistent design structures. If all of your proteins look identical, there are a lot of problems that this creates. But then people started adding more and more on top of this.

31:27Matthew McPartlon:Yeah, that is that's an interesting point that I think some people have acknowledged in the community. So how did you solve that? Yeah, you can see that if you sort of use your Oracle and also your sampler at the same time, you eventually will converge. What do you do to stop that or to convince yourselves that you're doing something valuable?

31:45Neil Patil:One of the nice things about structure prediction methods is that usually you have some calibration how kind of how confident the model is in its prediction. It turns out these models, they can give you a pretty well calibrated confidence prediction. So rather than just say, this is what I think the structure looks like, I'll say, this is what I think the structure looks like and kind of like, here are the parts that I'm not really certain about. and you can kind of aggregate this down to like a single scaler. And typically what people do is they'll look at like, okay, like not only how self-consistent am I, how much does this independent folding model even like the structure that it output?

32:19Neil Patil:So that was one way of early on, I'd say to like just gain confidence. And then like another thing that people often do is they'll look at like the diversity of their generations. Because again, you could have a model that's perfectly consistent, gives you great confidence predictions back, might be the same structure every time, like same sequence every time. So you also want to see like, okay, how diverse are the solutions? How many of these new problems can I solve in a sense? If I had a whole lot of money to validate, how would you do that? Can I go and, you know, do CryoEM or something like that and try to figure out the structure, you know, sort of get some ground truth on that?

32:56Neil Patil:It's more that the feedback loop is really slow. So you can validate a few structures like this, but it might take months. And it's just not like a very scalable direction. So I think that's like a problem for the field as a whole. And I think people are spending a lot of time, even like especially at CHI, I think thinking about how do we validate these problems at like bigger scale? How do we, you know, basically increase the throughput of our validation or increase the cycle time? Because if you're waiting months to figure out, hey, was my model correct? Like it's just, it's hard to iterate in a research environment that way.

33:24Matthew McPartlon:The good news is that this is getting a lot better, right? Like there's a whole network now of wet labs that you can work with that will run, you know, these assays, these experiments and tell you things about, say, you know, does your protein that you came up with bind to its target well? And so, you know, thankfully we're not at years, right? We're down to like weeks, which, you know, not as fast as like LLM land where you can just, you know, scale up and eval with and throw more compute and get results back in hours. But, you know, fast enough to where you can start to recursively self-improve.

33:52Matthew McPartlon:And, you know, I think we also spend a lot of time, like, you know, figuring out what are the metrics that we can compute, you know, in silico, like on the computer that are predictive, perhaps, of lab success. But, you know, your question about cryo-EM, yeah. I mean, also, you kind of have to measure the structure. And as you know, that's, like, so expensive because you have to kind of freeze the protein and shoot these electron beams at it and see how they bounce off. I remember there's, like, this really funny anecdote. We'll see if I can share it. But, like, the, you know, the paper in Chai 2, we actually, you know, did that.

34:20Matthew McPartlon:We took some of the proteins that the model predicted and ran cryo-EM, and we got the results back. And we were like, wait, the results look wrong because we had overlaid the kind of prediction over the point cloud. We didn't see any difference. And point being, we're getting the point now where these structure prediction models are within a few angstroms or less of the actual atomic positions that you validate.

34:45Neil Patil:In this case, it was a 0.33 angstrom error, which is one third the width of an atom. And we were like, this can't even be right. Clearly, they just sent us back the wrong design. They just sent us back our design. Yeah, exactly. Did you check for data leakage? Yeah, in this case, we actually chose these targets specifically to have no known antibody binder. So if we did get a hit, it was definitely the first antibody hit to this target.

35:12Matthew McPartlon:I think that's one of the things I didn't realize about biology was like just how much of it is literally feeling around in the dark. And that's not even a metaphor. You literally can't see like how these things look, right? So structure models are so, so huge because now you can, okay, you can actually predict within an atom, you know, how these things look. And that enables you to then do things like Chai 2 with the design models.

35:33Neil Patil:This to me is AI for science is one of the cornerstone problems, right? It is that you don't know, you fundamentally don't even know how to measure your problem in a lot of cases. So it's very difficult to validate. So you're getting these sub-Angstrom predictions with CHI 2, CHI 3, why CHI 3, what's better or what? Yeah, I think like with CHI 3, so like honestly, like there was a CHI 2, there's a CHI 2.5, there was a CHI 2.7, there was eventually a CHI 3. And like each time we saw better and better performance. and I think like the main thing with Chai 3 is like we look at Chai 2 and like we look at the targets it can solve there was like a lot of internal discussion after Chai 2 like hey we made like successful molecules binders to half of these 50 targets what about the other 25 you know what can we do to make those better and then like you know we were split we're like alright should we like study these targets that we missed and like figure out exactly like are there properties of these that we can look at or should we just bet on the models like will the models just get there if we put more time into like you know just be bitter less and pilled in that sense and just really bet on the models getting better and we we definitely took the the latter approach like we bet on the models getting better and we just pushed as hard as we could on that front scaling up the model the data whatever to to just build more accurate models yeah is accuracy is that the main thing is it binding affinity what do we so i think binding affinity is a big one like you you you can't just bind weekly in order for this to be like a useful tool, especially for our partners.

37:07Neil Patil:We need to start producing molecules that are like at or very close to therapeutic grade, which means like they have to bind really tight. They also have to be developable. They have to have like all of these nice therapeutic properties.

37:19Matthew McPartlon:And developability, I think that we talked about, he mentioned CHI 2.5, right, which we released like a few months after CHI 2. There was a study we did on the developability of the molecule, which, you know, for the audience, like obviously the molecule has to stick good and stick tightly, but, you know, there are these other properties you care about and to use the non-biological terms, right? Is it safe? Is it stable? Is it easy to manufacture? Does it, you know, self-aggregate? And we've been pleasantly surprised at, you know, how much we've been able to climb and push the performance in those areas.

37:50Matthew McPartlon:It seems like one of the reasons that you want to do antibodies is because the developability. Yeah, you get a lot for free there, right, with that antibody framework. Yeah.

37:59Neil Patil:It's interesting. I mean, to me, there are many structure prediction molecules out there. I mean, models out there. I feel like it's these other ancillary factors actually that are going to probably be the most impactful and the usefulness of a product. Yeah. Right. Yeah, absolutely. The nice thing about structure prediction is there is a ground truth that you can compare against. For design, you don't really have that. You're like, here's some new disease molecule. Give me a binder for that. And if you want to know if this thing really binds, you have to send it off to the lab and wait a while.

38:34Neil Patil:For structure prediction, you can be like, all right, the model hasn't seen this sequence before. It's never seen anything close. Does it actually like fold up into the correct shape? And we can just kind of hold that out of the data set and check. So I think I've always thought of structure prediction as this really nice speed run kind of benchmark to like validate ideas on. Right. Sorry, I didn't mean to say, I meant, you know, sort of structural models in general. Yeah. But yes, exactly. So maybe we can talk a little bit more about start getting into the product side of things. Thank you for coming.

39:06Neil Patil:I actually, I mean, like I said, I really think this goes throughout not only for, you know, sort of structural models like this, but also virtual cell and whatever. It's really all the other stuff around the direct development process that is going to have the biggest impact. So you talk a little bit about that?

39:25Matthew McPartlon:Yeah, I think that's actually a good thing to talk about after Chai 2 because I think Chai 2 is where it's started to get really fun from a product perspective, right? I think with Chai 2, we crossed the threshold of usefulness where after we released that paper, we had a lot of pharmas and biotechs approach us and say, hey, this model might be able to do some stuff for us. Can we use it? And then we're like, oh, man, we should build a product. We should build something to let you use that model. And that's right around when I joined. And there was sort of this, you know, mad, mad build out to both, you know, build the product, which we can talk about the shape of, and also go and secure the compute, actually, so we can go and serve those models to our partners.

40:04Matthew McPartlon:And, you know, I think another third piece there that was really interesting is, you know, around security and IP, right? I think we want to be a very neutral platform that anyone can design medicines on. But as you guys know, like pharma is this notoriously IP sensitive industry, right? And I think when I joined, a lot of people told me this can't be done. And like, they're not going to put their data in a platform and like have all their new medicines be generating out of it. And having a bit of a background in security helped a bit. Whereas like, no, actually, if you like just are really aggressive about how you like segment data and set up like single tenancy where you're like almost deploying a separate version or separate account in the product per customer, you can actually like build a platform and then go and ship it to them.

40:44Matthew McPartlon:And so, you know, through the summer of last year, we started doing that. Right. And, you know, we'd been working with, you know, or talking to Eli Lilly and, you know, they were, you know, one of the first partners to really work with us closely on that. Kind of, you know, it made that V1 of that design suite, right, that you can use to engineer some of those molecules on. And, you know, maybe it's worth talking a bit about that design suite, right? I think we have these really, really powerful models now that can do all of these crazy things if you condition them in the right way. If you kind of give them the right context about the structure that you're going after or maybe the constraints around the model.

41:26Matthew McPartlon:Like, hey, I want to design an antibody that hits this GPCR protein but doesn't collide with the cell membrane and also targets the specific epitope on that as well. And we looked at it and we're like, I guess we could put a chatbot around it. That'd be like really easy to talk to. But like really like you're trying to build something almost very visual, right? And you can finally build something really visual with some of these structure prediction models. And so if you kind of look at the CHI product, it looks a lot less like, you know, a chat GPT and a lot more like Autodesk or SolidWorks or Figma, you know, if you've used those things where you can kind of load up your molecule.

42:02Matthew McPartlon:There's this almost like Photoshop-esque like design suite. You have this equivalent of a paint tool to kind of paint your epitope. You have this clone of a content-aware fill tool to kind of get your binders generated from chai. You, of course, have a lot of the scientific analysis and plotting and whatever to understand the results of the models. But we've just been surprised at, like, how much complexity is actually just in, like, doing that right so that you kind of don't shoot yourself in the foot when you're then prompting these models to give you binders.

42:30Neil Patil:So are you sitting with people who are designing these antibodies, you know, and feel like they're complaining to you or whatever? Yeah.

42:40Matthew McPartlon:How do you convince med chemists to use your tools? Because med chemists hate AI tools. Like, notorious, like, I don't want to touch this thing. Or like, I don't understand it. And they will not touch things, which they do not understand. Well, it helps a lot to have the models working really well. Right. So when we, you know, when we had the results of CHI 2 and CHI 2.5, I think, you know, that's enough of an activation energy where, you know, pharma companies and scientists within these companies are like, oh, let's try it. Actually, can you, CHI, can you guys just try running the model against a few of these targets and let's look at the results?

43:13Matthew McPartlon:And then we do that and the results are good. And they're like, OK, let me let me try to get on that product and let me try to use it.

43:18Neil Patil:No, I think pharma is like incredibly pragmatic, actually. Like I've been very impressed with everyone that we've worked with so far. They're very, like I was saying, pragmatic about this. And they're like, they're willing to be proven wrong. And like, I actually don't blame them for not trusting the models. Like I have used these models. They've been burned so many times. They've like rightly so. Like I am pretty skeptical when I like see any release. I always have been. So like you really just like need to show them the proof. And like they can give you this target that they are interested in.

43:48Neil Patil:or maybe it's more of something they've worked on in the past. They probably don't want to like share IP right out of the gate, but they can be like, hey, you know, I've had trouble with this particular target in the past. Let's see how you guys can do on this. And then once you show them the proof, they like almost overwhelmingly are willing to accept that.

44:03Matthew McPartlon:I come from a cybersecurity background or, you know, have worked on security products before. And those were dark, dark years because you spend a lot of your time actually selling to people who are surprisingly not that technical. You think cybersecurity people are very technical. In many cases, they are not. And it is this kind of like uphill enterprise slog to this very unsophisticated customer. I think we've been just pleasantly surprised, or I have, by just how much I enjoy working with our partners and our customers. You know, these are scientists who have been spending, you know, 5, 10, 20 years of their life working on one target, right, often in some cases.

44:36Matthew McPartlon:And they've studied everything about it. You know, they're very sophisticated. They're very smart, right? You know, getting to collaborate with them is just a goldmine. And we learn a lot about how to make the product better. You know, there's this anecdote. We, you know, a few months ago, we were actually showing some of the results that we, from a target data with a pharma partnership. And one of the scientists in the room, like, started tearing up and crying. Oh, wow. You really hit that one. And she was like, and we were like, what's wrong? And she's like, no, I've just been, I've literally spent 10 years trying to get an initial binder to this thing.

45:13Matthew McPartlon:And you guys were able to help me do it. Oh, that's awesome. And, you know, that feels really special. To answer your question, you know, we, you know, there's, of course, the teams of scientists and computational biologists that we're working with within, you know, each of our partnerships. There's also the people we have within the building, right? So we, I think one of the things that I really appreciate about CHI is how cross-disciplinary it is. Like, you know, we have people who are maybe engineering experts and less bio experts like myself. We have great, you know, AI scientists, but, or ML scientists, but we also have a bunch of scientists that we work with and have joined CHI to sort of help us both, you know, test the limits of the models, right?

45:51Matthew McPartlon:See what is CHI-2 actually capable of? What targets can it do? What can't it? Inform some of the research direction there.

45:57Neil Patil:I want to add to that, like, in like the CHI-2 days, like we kind of started with like a bunch of engineers and people who have like AI bio experience. We didn't have a hardcore lab scientist. And like one of our first hires on that realm was Nathan Rollins, who I think he started working in the Baker lab at 14, graduated from Harvard at like 18 and got his Ph.D. by like 21 or something like this in the Marx lab. And he was like super skeptical about Chai at first. And then, you know, the results start to come in. He's like, OK, this is this is kind of interesting. Like this could work. And then like once the Chai T results came back, he was like, I need to bulletproof this.

46:34Neil Patil:Like nobody celebrate yet. So I think it's been really nice to have that level of rigor and to just have people who have really, they've spent the time in the lab, they've designed proteins themselves. They've literally, in the case of Andy, led several therapeutic programs, brought drugs to the clinic themselves. And we have all these people internally at CHI just using the products and really battle testing that.

46:56Matthew McPartlon:So if you don't have your own platforms, right? So you don't have your own programs, right? You're a pure platform or partnership model, right? Yeah. How do you battle test something if you basically aren't, you don't have a use case where you have to continuously push it forward? Or if you are just pushing things forward, when you just end up with your own candidates, if you're successful, and then what do you do about that? I mean, we have benchmarks of our own internal cases, right? You know, there's a set of targets that, you know, have our known therapeutics, right, that have known therapeutics against them.

47:26Matthew McPartlon:There's a set of targets that we pick to sort of push ourselves, right? And so we're constantly refining that set and adding to it. And that's what that internal science team that we have helps with, right, is expanding that and almost running the experiments to try to get initial binders there. We don't care about going and developing those drugs. Like, we just do that in service of validating and making our models better. And then, of course, there's a loop with our partners, too. Would you consider yourself hit discovery or are you, I guess, using some jargon, hit to lead, lead optimization?

47:55Matthew McPartlon:Like, where do you live in this? And, you know, hit discovery might be like one part of it, which you can do hit discovery. but the later, the other parts of this are, I think, oftentimes much more bespoke and kind of special. I mean, how do you balance that? And it seems much more difficult to me to be general than it does to solve general lead optimization than it does to solve, like, hit discovery.

48:19Neil Patil:I think ideally, like, we really want to be able to, rather than think of this as a bunch of stages, I think part of the reason why we think of it that way is because the initial molecules are usually, like, not good enough to be drugs. And like really like we're kind of at the inflection point now. We're really seeing this internally at CHI where the models are getting pretty close to like producing molecules that could eventually are like are very close to drugs. So we try not to make too much of a distinction between, OK, hit discovery, lead optimization, all of the different parts of this kind of preclinical pipeline are like, you know, the light.

48:55Neil Patil:The North Star is to just really produce drug like molecules straight out of the models. Of course, this is going to be hard. And like, they're going to be like tons of roadblocks. And like, you need to be able to like actually prompt the model to do this. You need the whole RL stack to like learn different properties, things along those lines. But I think it's very achievable.

49:12Matthew McPartlon:Yeah. And I think to add to that, right, yeah, this notion of target discovery and hit discovery and optimization where each of these has a gate and takes a few months to a few years is this very like waterfall model, right, where the cost of trying things and getting things early is very expensive. But I think to what Matt's saying, right, if you start to get in a regime where you can have models give you really promising candidates, you can start to make that look a lot more like a loop, right? It's akin to, like, becoming more agile in software development. Internally, we kind of have two, you know, North Stars, right?

49:43Matthew McPartlon:And at first pass, they almost sound like contradictory, but, you know, the North Star in research is to start to de novo one shot, you know, better and better and better medicinal candidates that are as close to being ready for, you know, the next phase as possible. But, you know, also within product, we do want to sort of expand into whatever these iterative workflows look like, right? Where maybe I get a binder, I get some results from the lab, I'm using that to condition my next run of the model. And I think, you know, they sound contradictory, but I think they're actually not. Because I think what's going to happen, you know, the research is going to get better at identifying a de novo candidate for like a specific class of drugs, right?

50:21Matthew McPartlon:Say like anti-agonists, right? Like blocking things, right? A little bit easier maybe. Okay, we can get to a state where we can one-shot pretty good drugs there. But now the next problem is like agonists, right? How do you reliably one-shot hitting a switch like on a cell, right? Or bispecifics or ADCs, right? And I think there's kind of this levels of abstraction that we're going to have to climb with the product as the models get better. One of the things I got very existential a few months ago because I was like, man, all this stuff we're building in the product to visualize molecules and do this.

50:54Matthew McPartlon:Maybe I'm just going to have to throw it all away when Matt ships CHI-4. But I think that's kind of the reality of building products now. You're actually using them less as an end in and of itself. Like maybe you'd have built software that was supposed to last like 20 years. Now it's supposed to last maybe one year, but it is the bridge to deliver value and kind of enable the research that then gets you to the next thing. And so I'd imagine we're probably going to rewrite our products at higher and higher levels of abstraction, right? Like maybe like right now we have something a little bit more akin to cursor where you're, you know, inspecting the molecule in the same way you're inspecting the code because you really need to verify like the bonds that are forming and the properties of the things that you're getting.

51:33Matthew McPartlon:but then you know you get to a point where that stuff is solved enough where now the product is actually just helping you orchestrate these like campaigns of hypotheses right or maybe you have like one target and you're like orchestrating a bunch of different epitope choices or whatever against that and then maybe you're going up one level of abstraction where you're now doing a whole campaign against all of the uh targets within a pathway right um and uh i think what's really exciting about that is if you if you have like these really good primitives for structure prediction and binding and design, and you can kind of compose them, then you can start to just like grow into like the outer loop of science, right?

52:08Matthew McPartlon:And then, you know, maybe the thing runs itself and you start to really get to some really, really, really cool drugs at the end of it. I actually want to push on what you just said about epitope prediction, because I think a lot of people in the field would argue this might be the much harder problem than finding antibodies and binders. Where do you think that the state of the art is in general and also with regards to chai in terms of epitope prediction and like is this a problem which has a reasonable solvable time horizon oh and also maybe can you define epitope prediction i'll think of this at like

52:38Neil Patil:some different levels so the most basic level is okay i have some disease that i want to target and what proteins are actually responsible they're like actually figuring out biologically what's going on like what should i be targeting in the first place with the drug because once you figure that out um it's kind of like a structural biology problem at that point you're like all right this like set of proteins is responsible and like what's going on there well this is interacting with some other protein that it shouldn't be interacting with. And conventionally, you'd just want to block that interaction or something with anybody.

53:05Neil Patil:But that's kind of where these proteins interact and the type of interactions that you want to disrupt. That's typically the epitope. It's the actual site on the protein that you want to block. This is a ridiculously hard problem. I'm with you on this. This is the harder problem. Just the amount of context that you need and the global understanding that you need to get in order to actually figure out what's interacting and how.

53:28Matthew McPartlon:But maybe let's take a few specific cases. Let's think about, what about SARS-CoV-3 comes around or the new flu or whatever. What would you do there? I mean, is that something that you think you could actually reasonably tackle?

53:40Neil Patil:In that case, like, yeah, you could just run a structured prediction model maybe and like see where the model thinks this thing will bind. If it's highly confident in that, you might say, okay, here is like the site that we want to block. I think in general, still very hard. And even like structure prediction, it's getting really good. And like a lot of people think, alpha-fold 2 like solve structure prediction not really like alpha-fold 2 got like i think 11 percent the multimer version of this got like 11 of antibody antigen prediction cases correct

54:08Matthew McPartlon:that means 90 of the time it's wrong yeah i mean but but uh alpha-fold 2 solved a certain class of monomeric proteins with msa absolutely yeah yeah so i mean the and that's the msa i think might be the key point here because msa's are sort of the the magic which makes it all work it's like a It's a template in some sense about what the structure should be. And antibodies almost evolutionarily can't have a template, right? Everyone has to have unique antibodies custom to the things that they've experienced over the course of their life.

54:40Neil Patil:And just to clarify, I had to understand this myself, so maybe I can help the listeners who aren't familiar. An antibody, the whole point of an antibody is it can be used by the immune system to identify new things that the body hasn't encountered before. So the design of antibodies as opposed to other types of proteins is to, the system is designed so that you can quickly recombine different components of it in order to match proteins that are from unknown pathogens, more or less. And so this is why it's not conserved in evolution the way that other proteins are. Yeah, so like back to the epitope prediction problem, I think it's still hard.

55:23Neil Patil:I think like there are a lot of cases that are maybe tractable, but I think in general, like if you want to discover this for a new target, still a really difficult problem. Maybe virtual cell would be like the closest thing to state of the art there, but that's still a ways out. I wanted to dig in a little bit on the product because there's something I don't understand about the economics of basically all the structural stuff that's happening right now. And obviously, a lot of people think it's very, very valuable. So, you know, I'm not grokking something. But when you look at the cost of developing an antibody, you know, it maybe is a couple million dollars, right?

56:01Neil Patil:when you go from, you've identified a target somehow, and then you say, okay, I need an antibody to match this, and then I have to sort of optimize it in various ways, and then maybe I try it in, I mean, with antibodies, you go to the animal typically faster. If you look at how much does it cost to bring, if you like are pressing it and pick the right target and the right technology to get all the way to drug, it might be half a billion. typically that$2.6 billion number is amortized over all the failures as well. So if you look at just the cost of that one success, depending on the disease, maybe less, but half a billion might be a good median number or something.

56:42Neil Patil:So you're saving like a couple million dollars in a half billion dollar campaign. So why is this so attractive?

56:50Matthew McPartlon:I would maybe challenge the premise a bit, like in a few ways, right? Like, okay, sure, if you're trying to get an antibody for like a very simple kind of target, like maybe, right? But I think what we've been most excited by is our partners using antibodies in, you know, more sophisticated ways, right? Like in, for example, in CHI 2, we showed like GPCR agonist activity, right? Where you can really hit the switch on a, you know, on a cell doorbell protein, so to speak, right? In a very precise way. Very, very, very hard to do that with antibodies if you can't be that precise, right? So you're unlocking a new capability.

57:22Matthew McPartlon:Yes. I would think about it as less like, oh, I'm taking the existing drugs that I can do and making them faster. I mean, there is some of that too, right? But it's like, no, they're just like, hey, how do you go after like better targets, right? That are, you know, maybe more precise, more effective, right? I see.

57:37Neil Patil:I think like also on top of that too is like there are drug modalities that you just can't discover with immunization. Like you're not going to design your like crazy multi-specific warheaded super intense formats. These are really things where you kind of have to design these from first principles. Even just with bi-specifics in particular like both arms need to now bind different targets. And you've kind of like have this multiplicative effect on your binding rate. So like if you have a one in a billion chance of finding a binder in ARM1 and a one in a billion chance in ARM2, this isn't going to work with a traditional approach.

58:12Neil Patil:Exactly.

58:13Matthew McPartlon:I think the other thing I'd think about is, right, you're not just helping your partner with maybe one drug, right? There might be a portfolio of targets that are there going after, a portfolio of drugs that they're trying to make. And the nice thing about the platform approach, rather than that we are developing individual drugs, is we can sort of scale with them as they pursue more targets in addition to more ambitious targets.

58:33Neil Patil:So, it lets you concentrate your learning, you know, subdomain of that and so that everybody benefits from that. Exactly. But, okay, so what are some of these capabilities you mentioned a few? Are there more that are really interesting that you guys are chasing?

58:50Matthew McPartlon:Yes. I mean, we talked about like, you know, cross reactivity. We talked about selectivity. We talked about some of these like really interesting additional modalities with bispecifics, right? There's a set of things that, you know, our partners have been asking us for that we've been working on that I can't get too into because then that starts to reveal some of the targets that they're going after. But I think the point being, you can just, once you get precise, like you can start to do some really, really cool drugs.

59:16Neil Patil:It's a new technology, right? So like technology in pharma means like how do you deliver your therapeutic? And so this is maybe kind of thinking about like CAR-T is a technology, right? And so this is maybe a new technology in the sense that you can have these highly engineered.

59:35Matthew McPartlon:Right. And that comes from the mission of the company is to really turn drug discovery from a scientific experiment to an engineering discipline, right? How do you sort of get to the precision engineering phase? for biology, where you can start with, you know, almost declaratively define the thing you're trying to get and have the model fill in the gaps and get you that. So what is the biggest blocker from going from science to engineering? Oh, man, there's so many things. Like, that's the thing about, you know, endogenicity. Yeah. Microheterogenicity. What is that?

1:00:10Neil Patil:I don't even want to talk about this. Like, the amount of headaches.

1:00:13Matthew McPartlon:Too late.

1:00:14Neil Patil:You already brought it up. You're kidding me now. Okay, so just like when you're actually parsing, first of all, file formats for biologists, they just don't care. There's no standardized, there are standardized file formats. Are they the best? I don't really know. But there's also just a lot of information that you want to pack. I have this structure. Here are the people who solved it. This is the method I used to solve it. There's a lot of stuff going on. And then depending on the method that you used to actually figure out what this 3D structure is, you might have multiple copies of that structure.

1:00:42Neil Patil:Part of it might not have really been resolved. where you're like, it could be here, it could be there. I'm just going to give you like both options. So like the actual just parsing problem on the engineering side of like working with this type of data is like really difficult.

1:00:53Matthew McPartlon:This seems like something that LLMs can excel at though. They don't know all the edge cases often, right?

1:00:59Neil Patil:This is more back to just like a simplicity approach. Like LLMs are very good. I will absolutely give you that. Then you're thinking about like, do I really want to like, should this function have 20 special cases or should we be like really principled in how we approach this? And should we be, I guess, more of a opinionated? Opinionated, yes. Like how opinionated should we be in how we do this? We want a strategy that's easy enough for humans to understand. And like when we're reading through the code base, we really need to know what's going on here. What are the potential problems? And like sometimes that just comes down to looking at examples.

1:01:31Neil Patil:But then I think, okay, once you've kind of figured out all the infra work and how you get data into the models, there's then like scaling the model. There's then scaling the infrastructure around the model to train bigger and bigger versions of this. And that's like a lot of work that Neil and the product team actually leaves.

1:01:45Matthew McPartlon:Yeah, I mean, that would have been my answer is the infrastructure part. I mean, you know, not to beat a dead horse, but compute, right? Getting the compute and using it in the right way is such a challenge, especially for startups. This has been such a theme. Yeah, yeah. Anthropocic is holding back science. It's not going to be higher. No, I mean, and to that point, like we... I mean, they're also accelerating science, but it's like this weird... No, totally. Totally. Like one of the things that I help a lot with at Chai is buying compute for the company. Worst job, man. I would not recommend it.

1:02:19Matthew McPartlon:It is very stressful. But, you know, even September of last year.

1:02:23Neil Patil:You're back to the hardware job. Yeah, I know. Exactly.

1:02:26Matthew McPartlon:In the wrong way. But, you know, September of last year, we started to really notice like things are getting tight. Right, right. We were doing a lot of our inference on, you know, spot and on-demand markets. And we'd have these days where you just like get these capacity crunches and we're like, okay, we should probably start to get ahead of buying some compute for ourself. And I mean, I think everyone probably says this, but man, it was hard. Like, I think I didn't realize how much of a power law, you know, this is, right, where, you know, there's, say, 10 ,000, you know, B300 units that are shipping everywhere, right?

1:03:00Matthew McPartlon:The hyperscalers and the, you know, the biggest AI labs are buying 95 plus percent of it, right? And then you kind of have the startups, like, fighting over the scraps. And I think the other thing that's really interesting, especially if you look at these later computer versions, right, the Vera Rubens or, you know, the B300s. like a lot of this stuff has been built very like LLM for it, right? Like you have these, you know, systems with like huge KV caches where you have like 72 GPUs that are all acquired to talk to each other, right? And, you know, obviously some performance gains there like help us, right?

1:03:33Matthew McPartlon:But like it's kind of interesting just how much the compute market has kind of gotten LLM pilled. I think there's like a whole probably set of, you know, compute stack and inference optimizations and things that need to be made for this class of models. And, you know, I think this class of models is going to be like just as big, just as impactful as LLMs. But it's almost like the compute market like kind of doesn't realize that yet, both in the capacity sense, but also in like the software stack sense. So we actually spend a lot of our time, you know, even just like doing basic optimizations of compute to like get them to work better for the types of models that we have.

1:04:07Neil Patil:Yeah, I know that some structured models are more recursive than LLMs, for example, which changes sort of like maybe the compute to memory ratio that you need and things like that. What are some of the sort of cool or interesting optimizations that you've done there? Depending on the type of model. So we can go back to a Chaiwan type model. In that case, we're following the LFOLD 2.3 architecture. and there you're like rather than doing attention over like this like normal sequence representation you're in a sense loosely doing attention over this pair representation so you can think of this as like a sequence of length l squared uh rather than like typically length l if you're doing attention over that the way that you actually batch this up it ends up being l cubed now you're you're in like a pretty pretty heavy compute regime uh so the amount of flops that you're putting into every token stays it's pretty high the amount of memory that like the memory bandwidth overhead of just transferring that from SRAM to whatever, that's a real bottleneck in these architectures.

1:05:11Neil Patil:So even something as simple as a layer norm can take a long time, actually. That can be a significant amount of the compute that you're using. So I think on our side, we've spent a lot of time just optimizing and engineering, taking engineering very seriously so that these operations are at least better. We're always looking at how do new chips perform compared to the older versions. sometimes that's even different for training versus inference. And like, of course, Neil knows this really well.

1:05:37Matthew McPartlon:Well, so there's, you know, what you're doing on the individual GPU. And then there's like, how do you like orchestrate fleets of GPUs, right? And, you know, you basically shard your computation, right? And so, you know, when you're designing a molecule on Chai, it's not necessarily like one call, right? It's a lot of GPUs being thrown at the problem, right? Across a lot of compute. And actually, I would say that one of the hardest things to get right in software engineering is durable execution. Are you all familiar with that term? Can I go on a little? Yeah, sure, go first. Ultimately, if you're like computing a lot of data, you know, model calls across like a very wide set of infrastructure, you always run into these problems where like some part of the infrastructure is flaky, right?

1:06:19Matthew McPartlon:Like maybe the bucket you're grabbing your data from like goes down or like your database has a blip because there are like too many transactions against it or you're like GPU errors out, right? I've been at companies before where you like spend so much of your time just dealing with this shit, right? Like you're basically putting like all of these cues and like all of these retries and you're like duct taping things together. And you have a and it becomes this mess where now what used to be like a idea, like a pretty simple computation that's just distributed. You're ending up spending like 95 plus percent of your time on all of this queuing and retry stuff, right?

1:06:54Matthew McPartlon:We're huge fans of this company called Temporal. Basically, you know, there's this idea, like, look, if you're just trying to get something, a really long running job to run, at the end of the day, what do you need? You need a queue. You know, you need your flaky thing, like, pulling off of the queue. You need some retry logic to put things back on the queue if they fail, right? And then you need some whole, like, orchestration system to just, like, tie all the queues together and monitor them. What's really cool about Temporal is, like, this is a tech, a company that's kind of invented a framework for doing this.

1:07:24Matthew McPartlon:And I think one of the technical decisions we made early on that was very helpful was to run as much stuff as we can on Temporal, right? So whether those are, you know, calls out to the database from the app, right, to make sure the database transaction goes through without failing. Okay, let's have side effects like sit on Temporal so that they get retried smartly without us having to like write our own queue logic, right? Or things related to model calls or things related to orchestrating really long data pipelines. Point being, like, you know, one of those primitives, like, just like, hey, you need to get durable execution right so that you're not stuck in, like, retry hell.

1:07:59Matthew McPartlon:A really deep, like, engineering thing that, like, you wouldn't realize if, unless you, for, like, me and Jack, you've been, like, burned by this, like, many, many times before. And I think, like, we're at this state now, right, where we've, you know, we've raised another$400 million. I have to go buy another compute cluster. Or like, you know, like we're going to have like really, really, really large runs and inference and training sets. And so getting those foundations right is what's actually going to let us do more ambitious things. And to kind of answer your question, actually, that's a lot of the bottleneck to making biology more like engineering.

1:08:33Matthew McPartlon:It's just like having the right engineering primitives supporting it.

1:08:36Neil Patil:I have an analogous tangent on the model side. Actually, one of the things that's kind of nice about those problems is they're like super visible. So at least you know, like, hey, this crashed, this failed. For us, we just see like, loss curve didn't go down or like we see weird gradient behavior or whatever. I think a lot of these same principles like engineering first, that also applies on the research team. One thing that I like to say is kind of like complexity and being bitter lesson pill, they're like fundamentally at odds. For example, I think like outfold three, I might get this number wrong, but I think it was like 23 sub modules.

1:09:08Neil Patil:And at that point, that's a really difficult system to optimize and study. You're like, all right, what happens if I change, like if I tweak this thing in submodule 30 or like 21, what happens to the whole system? And you can always think, hey, we can make this better by like adding module 24. But like, should you or should you think about just like removing things and lowering that complexity down? But I think that's like a pretty fundamental thing at Chai is just like the engineering culture and just being like very simplicity biased.

1:09:33Matthew McPartlon:Have you all seen the picture of like the SpaceX engines? It's like Raptor 1. Oh, yeah. It has a bunch of pipes and like Raptor 2. We have a picture of that like on our office wall because I mean, it's just true, right? Like how do you delete, delete, delete more things? Yeah. But the only way you can accomplish that is, I mean, the reason AlphaFold 2 and AlphaFold 3 worked, they were small models, relatively speaking. They were very compute intensive, but they were very data efficient. Yes. And like there was inductive bias after inductive bias brought in by human intuition and probably like hard fought experience.

1:10:09Matthew McPartlon:it was they're incredibly efficient if you try to knock down those things you know they're not like a house of cards like everything is a incremental improvement on top of it in order to get beyond that it seems to me like you really just need new sources of data you need to at least treat data fundamentally different in a way that is much more efficient I mean I mean I'm actually kind of surprised to hear that you have scale to that degree because it suggests that you're doing something very different from what the community is thinking, the way the community is thinking about it. I don't know if you can comment about that, but.

1:10:44Neil Patil:We're pretty first principle people. Like the whole research team at CHI, except for me and Kevin really, like we're the only people with quote bio background. Even still, like we're pretty far removed. So I think like we try to like look at every problem as a core ML problem. We try to think of like what's the analog in other spaces. So like even for image models like cnn's were built to process images so like images should be looked at in patches like that was the nice inductive bias there and then people were like well you can just kind of tokenize this thing throw it into transform and it's going to work and like it did end up working uh even like on a relatively small data set but i think um for proteins in particular it is really hard there's not as much structural data there's a ton of sequence data and like that's one of the unlocks for like esm working uh you can get that to just run on a transformer If you try to do the same thing with like experimental structure data, good luck.

1:11:34Neil Patil:You need AlphaFold. Yeah, absolutely.

1:11:36Matthew McPartlon:There was that Apple paper where they distilled on the AlphaFold, which it was actually really cool that you could distill on a very large data set and you could get, you know, good signal. But, you know, it didn't generalize at all because it wasn't reasoning. It was really pattern matching. Like one of the things, these like triangle layers you were talking about, for example, they do have a very nice inductive bias. Maybe it's not the triangle inequality like the paper originally proposed, But it's a clean, adductive bias, and it unambiguously is like one of the things which made it work. And it just comes at a huge cost.

1:12:08Neil Patil:Yeah. Yeah, no, I think that's definitely true. These layers are pretty costly. And like that kind of limits what you can do with the architectures. They're not like, not only are they like costly in terms of compute, they're just like not efficient on modern GPUs either. You have small hidden dimensions, large sequence dimensions. Like it's like exactly the opposite of what GPUs are designed to process. us one takeaway from like triangle layers is you're kind of just trading off parameters for compute in that sense like that's like one mental model for thinking about this uh i might want to like throw more compute at the problem and just trade that off for parameters because like i won't be able to hold as many like i can't literally store these you know large pair representations and still do normal attention uh so i think there are fundamental things you can abstract from the ideas like alpha fold um but you can kind of just like tweak these and start billing off of them in your own way.

1:12:57Matthew McPartlon:It sounds like you have quite a bit of research, like fundamental research going into this direction for, I guess, audience looking for a nerd snipe and ML engineering for new problems. Probably something very, it's a very different research direction than a lot of the communities going in.

1:13:14Neil Patil:Yeah, yeah. I think what we built at Chai is like, it's very unique in a lot of ways, but also very tied to like what CoreML is good at. Kind of what I was saying before, Like we try to map every problem into like a core ML problem. We think, you know, how would you approach this if it were an LLM or something like that? But yeah, like at the end of the day, we really, really value simplicity. And we really encourage people who don't have a bio background to like not be scared of this stuff.

1:13:40Matthew McPartlon:And I think that extends into the product too, where, you know, there's a balance to be had here, right? Between like how general do you make the product? Like, do you build a cross-reactivity workflow and a selectivity workflow and a bispecifics workflow? Or do you all say, no, like, let's make the model general enough to say I'm going to, like, condition on arbitrarily binding or avoiding something. And then you just have a very general, like, screen in your CAD suite where you can say, hey, I just want to avoid or bind to these parts of these different structures, right? And I think, you know, kind of like the ML team, like, I don't have, you know, a formal bio background.

1:14:14Matthew McPartlon:Most of the product and platform team doesn't have a formal background either. Now, there's some amount of like maybe regretting my words that I'm going to have, right? Because I'm sure there are, you know, a million nuances and, you know, I don't want to come off as, you know, too brash or naive there. But, you know, I think sometimes it's helpful to not be burdened by like all of the, oh, this nuance and this nuance and this nuance. And you get to kind of bet and be maximally general because, you know, that's kind of what we're seeing in the research. You can, the models are very general. That lets the product be very general.

1:14:43Neil Patil:I'm thinking back to like in my in my CS theory days my first advisor was like we're working on some problem and we we need like a polynomial time algorithm for something and he would always tell me like never underestimate the power of polynomial time like this is basically like you're allowed to choose like whatever exponent you want and my first paper was an n to the 20th time algorithm for this problem and I was like Andy I did exactly what you said he's like wait a minute I didn't need it like that yeah I think You can really help yourself. You can free yourself up a lot when you're like, all right, I can kind of do whatever I want and then kind of simplify it later.

1:15:16Neil Patil:And I think that's really a pretty fundamental way of thinking about things that we leverage a lot at CHI.

1:15:23Matthew McPartlon:The space of protein design and binders in general is actually a fairly crowded space. I'm curious about what your general outlook of the field, the industry is. I mean, I can go back to some anecdote. I was at maybe NeurIPS three, four years ago, the one right after RF diffusion came out. I was talking to someone in the Baker lab and they're like, man, I just one-shotted. I don't think they even used one-shot. One-shot wasn't even a term back then, but they was like, I just got picomolar binders out of RF diffusion and just like threw in the cryo. Great, right? It didn't seem like that just solved the problem.

1:15:59Matthew McPartlon:Like, it's not like, oh man, now everyone. Yeah, but there are lots of people who I think have seen that you can actually do protein design, at least in some categories, quite well. I'd say like, is it mini proteins or mini binders? Ironically, nano binders are actually smaller than or larger than mini proteins or maybe like a little bit harder. Antibodies are typically considered even harder. But there's this like, is this something which can and will be commoditized, at least in some part? How do you compete? Like, where does this, where do you, where does the field go from here? I mean, I think the answer is it's kind of all of the above.

1:16:37Matthew McPartlon:Like, I think there probably will be some commodity layer for certain types of modalities or drugs, right? I think at the same time, we're going to be able to do even more and more and more ambitious drugs. And you're going to, it's just like what's happened in LLM land, right? Like you have your open source models that are maybe general and helpful for some things, but people are still buying frontier models, right? And actually, if you look at the amount of value captured, it's actually the closed source frontier models. The whole pie is growing, but it's growing so fast that even as the open source models like share expands, the frontier models are still able to capture the majority of the value.

1:17:13Matthew McPartlon:Raise your hand if you're using an open source model on your day to day. Right. And what are the reasons for that? One, like if you have more intelligence, you're going to go after harder tasks. Right. I think if we have more, you know, intelligent bio models, we're going to go after more, more crazy bio tasks. Right. But then also, too, like, I mean, a lot of the reason I don't use the open source model is because, like, you know, I don't get like cloud code. Right. I don't get like cloud. You know, I think there's a product layer to be built that is just as important as the model layer. We learn a lot from our partners and, you know, the people in the building as well.

1:17:46Matthew McPartlon:Just like what are the really tough things that they get stuck on using the models? right? And some of them are like, you know, the dumbest things, right? Like, you know, I want to be able to better visualize this piece and like focus on that. And some of them are actually like very sophisticated things that we then have to build some like pretty vertical product for. And look, maybe in the fullness of time, like AGI, like one shots everything and doesn't matter. But I think there's quite a bit of time until we get there, right? And I think the product makes a huge, huge difference for that. That'd be my answer.

1:18:16Matthew McPartlon:I mean, you probably have a more model forward answer?

1:18:18Neil Patil:No, like I think like like biology is slow, which is like one kind of nice thing. And there's like not that much labeled data. So like you could take all the publicly available sequence information out there that might give you a good base model, but you still need some measurements on that data. That's still pretty time consuming. And then you need to like iterate on that. So I think there are even just data blockers there and unlocking like if we really want to do this zero shot design candidate, start generating molecules that are almost ready to go into the clinic. I think it's more to that than just like, you know, AGI might not solve that right away.

1:18:52Neil Patil:I think there are definitely like some technical blockers there.

1:18:55Matthew McPartlon:But even in the space of, you know, specialist companies, I mean, I'm not going to like to start naming them, but there's, I think, I don't know, probably 10, 15 protein design startups. I think the two things which it sounds like Chai has gone on is like one, all in them product and two, you are not trying to do your own platform. If you don't have your own data mode, you know, Is that going to help you one out in the end, or is that going to be a blocker? I'm just curious about that.

1:19:22Neil Patil:Yeah, that's a great question. Yeah, so Chai, definitely no plans of starting a pipeline. We take the partnership model pretty seriously. And I just, from a personal stance, I love the incentive alignment, and just being like, we make the models better, the partners succeed more, and just that iterates on itself. So I think that's a pretty unique part of Chai, is like one, just being able to partner with a lot of people to getting like the feedback on the product. So like, you know, knowing that it's very real, this is in like, like legit big pharma hands and they're actually running campaigns on this stuff.

1:19:56Neil Patil:So I think it's interesting. We really have to be model forward, model focused. Like we need to keep delivering value. So that puts a lot of pressure like on the research team, the product team, first of all, to like to serve these things, the research teams always shoot for like better and better versions. The way I think about this is like, if you're a bitter lesson-tilled forward kind of like thinker or company, then there kind of comes a certain point where there's a lot to do on like both the model and data side. But I don't think either is exhausted. It would be stupid to say like, we don't need any more data, but it'd also be stupid to say like, the models are stuck.

1:20:30Neil Patil:We only can like use data to solve these problems. So I think there's like tons of room to grow on both sides. We're taking like both very seriously.

1:20:37Matthew McPartlon:And I would also maybe push back on the no data moat premise, right? That'd be kind of like saying, hey, like all the enterprises that work with Anthropic, like you're not letting Anthropic train on their data. So like you can't build models that are good at enterprise workflows. I think, you know, one, we are investing in this. You know, there are ways to turn compute into data and get more and we're doing those. But then also, too, OK, what is the kind of data that you're trying to get? And I think what is kind of cool about working so closely and supporting so many of these partners is we get to really learn about what is the stuff that would be helpful in research.

1:21:10Matthew McPartlon:And so rather than doing research in a vacuum based on what would hypothetically be cool, we're able to sort of kind of do informed research based on what our partners have just been very organically asking us for help with. I see. I assume that you aren't allowed to train general models based upon your partner's data. Do you train specialized models for like, is there a Novartis model and a Pfizer model? Yeah, I mean, like a lot of these deals, you know, and this is all public, right? We are working with them to, you know, train or fine tune a version of our model for them. And I think there's probably like so much more we can do there over time.

1:21:47Matthew McPartlon:My brother started a company called Applied Compute, a great company. They're kind of doing this thing for, you know, design for LLMs, right? in helping enterprises really understand the value of their language data and do that for specialized tasks. I think there's a whole world where we could potentially do that for biological data.

1:22:02Neil Patil:What is the value there? Like, what is the lift that you get from using their data? I mean, is it just that it's more data or is it more that it's specialized to a problem?

1:22:12Matthew McPartlon:You know, they have a lot of, like, scientific, you know, data that they're getting from experiments that can maybe help our models do better in, like, particular classes of candidates or targets that they care about. Yeah.

1:22:23Neil Patil:I mean, even something as simple as like they might just have some preferred way of doing things that might not be like native to the CHI model. And they can like, you know, kind of like ask the product team and in a sense to just be like, hey, we like, you know, our designs have property X. Can you make sure that they have those? So I think like even things as simple as that actually have like a pretty big impact for them. Yeah, so I mean, this goes along with a pet hypothesis I have that all AI companies, and especially bio and scientific ones, are actually consulting companies. Pharma, I think, is particularly the case because you're developing a new drug, right?

1:23:04Neil Patil:It's almost by definition new, right? So like the existing stuff has to be customized in many cases, right? Unless you're doing something that's just reiteration of old stuff. But a lot of the big pharma are pushing the boundaries of science.

1:23:17Matthew McPartlon:Yeah, I mean, certainly, like, we aim to make the models very general. We aim to make the product very general. We aim to make it powerful. But, yeah, I mean, there is integration work, right, with every customer. To answer your question, you do get some defensibility just by doing that, right? And I think what is nice about building, you know, trusted relationships with these partners is hopefully, you know, if we execute really well over the next, you know, the first year, then they'll continue working with Chai to do more ambitious and more drugs past that.

1:23:46Neil Patil:I mean, it's going to be hard to switch, right? I hope so, yeah. Just getting the security review done. Yeah, yeah. Maybe one other interesting point is if you think of this on a per token basis, I don't know if there's another domain where the downstream value of a token is as valuable as it is for pharma. Yeah, that makes sense. You're thinking about the actual drugs that come out. These can be multi-billion dollar assets. So like the case of GLP-1s, I think the two GLP-1 drugs combined are like maybe a trillion dollar asset.

1:24:14Matthew McPartlon:Yeah, I mean, I think up until I think three months ago, right, GLP-1s like total revenue was more than all of the AI labs put together. Yeah. I don't think people realize that. Like I didn't realize that. It's crazy, right? But yet the market cap way lower. It's like crazy how relatively speaking the market cap is. And, you know, I didn't realize how much of like a VC business, you know, you know, pharma is in, right? They're in some sense like taking really ambitious bets. You know, I think one of the things that was really cool with, you know, is like if you study the history of Silicon Valley, right?

1:24:44Matthew McPartlon:Like obviously people think of Silicon Valley with software, but, you know, in the 80s, one of the biggest venture outcomes, one of the first ones was Genentech, right? And because it is such a VC model, right? You get the string of tokens that can then give you so much value downstream. Just a general shout out to Outposting's blog series. It's about like finance and funding and yeah, really fantastic. Before that, I knew a lot of those points, but I did not realize just how deep that rabbit hole went. Yeah. Yeah, I mean, it's maybe the single biggest problem in biopharma is actually just the funding model.

1:25:23Neil Patil:There's also, have you heard of Aram's Law? Yeah. Oh, yeah. Yeah.

1:25:27Matthew McPartlon:You know, Moore backwards.

1:25:28Neil Patil:Yeah, Moore's Law backwards. So it's like, and like compute, you know, it's kind of scales. So you have like this nice exponential scaling, log linear scaling of compute, and you have the exact opposite in pharma. So like the cost of actually making a drug in pharma is kind of like increasing exponentially. So the amount of money put in per drug is growing at kind of like an exponential rate, which is, it's pretty interesting to see this, yeah.

1:25:50Matthew McPartlon:Which guarantees at some point, the marginal return on a new drug development will be negative. Exactly. So unless someone, maybe Chai, figures out how to, you know, fix this. I think that we might be on the verge of sort of flipping some of these. Phase transition, bending the S-curve.

1:26:09Neil Patil:Yeah. Just to double, maybe belabor the point, but that pharma and VC fundamentally both are optimizing a portfolio. Yeah. And I think that's the connection there.

1:26:19Matthew McPartlon:Yeah. Thinking of pharma as like sophisticated capital allocators, right, where they have this portfolio of targets and they're allocating between them, I think that was a big reframe for me. Yeah. And I think we will just see more of that in the future, right? And hopefully they can take, you know, in a sense, the VC taking riskier bets. Like hopefully pharma can take riskier bets and pursue really, really cool drug targets in the future.

1:26:40Neil Patil:That analogy is actually like one, the kind of like VC type investor-ish model. It's like actually how we think a lot about research at Chai as well. Our research team is relatively small. I think definitely compared to like a lot of the isomorphics, deep minds. Like our research team is like, you know, in the around 10 people. So, like, we're a relatively small team, but we kind of think of it as almost like an investing job where, like, you're investing ideas towards compute. In the same sense, you're really just capital allocators in that respect.

1:27:10Matthew McPartlon:Yeah, I actually think maybe this is too cute, but I would even make the broader point, which I think we kind of think of everyone at Chai as a bit of a capital allocator. So, I think one of the things that surprises people is we're pretty small. We're only 30 people. And that's because everyone we hire onto the research team or the engineering team, you know, especially now that they're in some ways like very empowered with AI, a lot of it is just like allocating, you know, their attention into the right ideas and allocating their compute.

1:27:34Neil Patil:This is actually, I think, a characteristic to some extent of machine learning AI projects and also science. Whereas if you're building like an API for some B2B SaaS company that's not building foundation models, whatever, your limit is mostly people. So the resource you're allocating is almost entirely people. Whereas if you're building hardware, you're building AI models, you're building something scientific, then your constraint is the resources that are, you know, the bottleneck is, you know, the lab, it's the compute, it's other things. And so that you have to really be in that mentality of, I have these limited allocation of, I have some shots on goal, how do I allocate those shots?

1:28:20Matthew McPartlon:Well, I would say yes and no. So I agree it's a bit more like that, right? But like, let's going back to the example of building an API for, you know, a B2B company, right? That API has incremental cost. You have to support it. It adds complexity to the product. It's another thing you have to go market and sell. Maybe you should actually be allocating that into like a different bet, right? A different thing on your product roadmap that you should be prioritizing instead of the other thing. I think in a world where like building things just gets like really cheap and, you know, increasingly free, the scarce thing is the attention both that you can put into it, right, to keep your product simple and grokkable and that your customer can put into it to like really understand how to use it.

1:28:58Matthew McPartlon:I see it less as like a binary thing and more just like we're all kind of as engineers going to be a little bit more like allocators of attention.

1:29:07Neil Patil:Which is what executives are. We're all going to speak coming.

1:29:12Matthew McPartlon:He says, you know, Microsoft wants to make everyone a manager of infinite minds, right? If you like really take that to your extreme, like everyone's going to be an executive. I mean, I certainly feel like an executive when I talk to Claude every day. A little suite of interns who are all going out and eagerly solving problems. You may or may not have actually wanted, but they're solving the problem. Yeah.

1:29:33Neil Patil:So we have two typical questions that we asked that we've already kind of asked one, but I'm going to ask it again, maybe more directly, is if you, and you can both answer this, if you could remove a bottleneck from your problem space by fiat, what would that be? That's an interesting question. I think one thing that'd be really nice, like just I'm like always in research land, very hard to turn off. For me, it's probably just the validation loop of protein design in general. So like just being able to say like instantly, like, hey, this thing works, this thing doesn't. There's still a bit of walking around in the dark that you're doing.

1:30:10Neil Patil:just to like, you know, you have some ways and like I think at CHI we've taken this like very seriously but it's probably along the lines of just like validating hypotheses and like, you know, knowing for certain that things work. Yeah, that's unsolved problem for sure. Unsolved problem, yeah. Yeah, and would be hugely valuable. Hugely valuable, yeah.

1:30:29Matthew McPartlon:I'm going to take a much more abstract answer to that which is actually like talent obscurity. I think, you know, there's a lot of smart people going and working on LLMs. You know, there's a lot of people that are working and becoming software engineers for SaaS, right? But I think just like not that many like smart people go and work on bio. You know, I didn't work on bio like in high school because I was like, oh, I could like pick up my computer and program apps. But if I want to work on bio, I have to like go study and get good grades in school and like maybe get a PhD or whatever, right? And, you know, maybe that's one reason for it.

1:30:58Matthew McPartlon:I think another reason is, you know, a lot of this stuff is really obscure, right? Like we threw around a lot of big words during this podcast. You can't really visualize the things. It's one of the things we care a lot about at Chai is like, how do we make the whole thing feel visual on our website and in the product? And, you know, part of the reason we're here is like, you know, I think, you know, more people should realize like you don't need to like have like a super, super, super specialist bio background to contribute to this like computationally. And so, you know, I think a lot about like talent flows and like where talent goes in the economy.

1:31:29Matthew McPartlon:And right, you know, in the 90s, everyone was flowing to talent. And, you know, since the 2000s, people have been flowing to tech, but, you know, big tech like ate up a lot of the talent, you know, until, you know, a few years ago. And now maybe like LLMs and the big AI labs are eating up a lot of the good talent. But it's like, you know, at the meta level, like how do you allocate talent better? You know, selfishly, I want more talent going into bio. I mean, we probably want more talent going into manufacturing and physical world things and these other problems that the U.S. has. But yeah, I think communicating that better would be the thing that if I had a megaphone to talk to everyone, I would try to do that.

1:32:05Neil Patil:Okay. So then that leads to the second question, which is, and maybe the answer is the same, but what is the takeaway, one takeaway that you would like people to have from the episode? Yeah.

1:32:18Matthew McPartlon:I mean, I think, you know, biology has been this somewhat obscure feeling field where you're stumbling around in the dark. You don't know what you're looking at. You're dealing with non-determinism in your experiments. You're having to do a very long and iterative trial and error loop across a very, very long amount of time. And at some point, you're crossing that threshold of what you can do computationally when you can get folding models down to being within, you know, an angstrom, right, where you can get design models to give you, you know, hit rates, you know, north of 50 percent, or now you can put them, you know, in a 96-well plate and actually have like 48 interesting binders, you start to get to the point where now you can declaratively precision engineer what you want rather than betting on, you know, nature or trial and error to get you there.

1:33:09Matthew McPartlon:And I think that, look, we had the same thing happen in software where you can write code and you can deterministically get an outcome or in electrical engineering where, you know, you instead of your schematic being drawn out, you can put it in cadence design systems and get it on, you know, it made in software, right? Or CAD for mechanical engineering, where you can sort of precision engineer your part and get it printed or manufactured. You know, the same thing is happening in bio and it's happening very quickly. And that really opens the door for a lot of really interesting people, or maybe it wasn't as escrutable or accessible before, right?

1:33:45Matthew McPartlon:Like software engineers like myself, researchers like Matt, you know, obviously we're still going to want the specialists, But, you know, the generalists can often really accelerate the precision engineering happening in the domain.

1:33:57Neil Patil:Yeah, I think for me, like, the base takeaway is that the field is actually working. And, like, not only does it have commercial traction, but, like, the research is, like, actually showing signs of life. Like, it's not even just showing signs of life. Like, the signs of life have been shown. We're actually in a place where, like, the models work. They're delivering value. And, like, there's still tons of really interesting research problems to solve. So I think there's a lot more low-hanging fruit in this field than there would be in other fields. And I think the amount of impact you can have, especially as a researcher, is just unmatched in this field.

1:34:28Neil Patil:For us, we're all very mission-driven. But even if you're not, it's a lot of fun puzzles to solve. There's this kind of 3D geometry angle. If you like diffusion models, there's a million problems to solve in that regard. We have this LLM-looking trunk in CHI 1. There's just so much of core machine learning is touched by these problems. We're still, although we've made a ton of progress, there's still a lot to be done. And I think it's just like one of the most interesting fields to be working in, which like while also having some of the largest impact on just like humanity. Cool. Thank you so much.

1:35:01Neil Patil:Thank you for having us. Thank you for making a long journey. Yeah. 22 minute walk. And, you know, we look forward to tracking Chai's progress. Awesome. Thank you guys. Thank you very much.

1:35:17a