Chai Discovery launches CHI-2, an antibody foundation model achieving 15-20% success rates and potentially eliminating high-throughput lab screening

Very cool. Well, I wouldn't be surprised if uh you start getting some major offers. I can see the the the AI researcher space is hinging heating up. The AI journalist uh it might be the next domino to fall. So yeah. Oh god.

We're here we're here to help talk you through any any sort of offers that you might get from you. You guys can take a percent. We'll work it out. Fantastic. Awesome. Well, great great having you on and uh yeah, excited to follow the rest of your reporting this week. Thanks. Cheers. Bye.

Next we have Joshua from Chai Discovery with a big launch this week. It's always tough to to launch when um the biggest trade deal in tech history is going down. Not even a trade deal, the biggest poaching. We keep calling it a trade deal. It was not much on the other side of the trade.

It was a non-consensual trade deal. It's a It's more of like a a draft and Zuck is just drafting from the entire league. I take 12 of your best people. Yeah. For nothing. Yeah, he's kind of just building an allar an all-star roster. It's like all-star weekend over at Meta this week.

Anyway, uh Joshua, welcome to the stream. How you doing? I'm doing great. Thanks so much for having me. Excited to be here. Of course. Great to have you. Uh kick us off. Great work, you know, uh launching on on such a a wild day in in tech history.

Um but uh is the time run really getting like completely steamrolled by this story or I don't know. It feels like that. Maybe maybe I feel like there's also people on vacation and stuff, but break it down for us. What is the big news?

Well, the big news is we've just announced CHI 2, which is our latest foundation model at at Chai Discovery. Chi 2 is actually allowing us to now generate antibbody sequences with a success rate that's high enough for us to skip the usual high throughput lab testing.

This means that you can go take a target, put that into our model, go generate a set of uh potential protein sequences, and then test them in the lab, and they tend to work with extremely high hit rates. Uh previous methods were like 0. 1% or so.

So, you'd have to screen, you know, hundreds of thousands uh to to potentially find something. Uh but here, the hit rates are actually upwards of of 15%, approaching 20% uh of success. Amazing. uh break down some immediate applications of that tech across maybe different types of uh categories.

You want to you want to ask him about making a more more potent testosterone that maybe we get to that next. Okay. Um yeah, may maybe that's the right spin output. I think one thing that we're uh really passionate about is going after areas that are really hard to do today.

So, you know, like this allows us to do things faster and cheaper potentially, but I think what's really going to be gamechanging is just opening up new kinds of biology that wasn't possible before.

So, we've got scientists that are like, you know, maybe working for years trying to enable a certain class of drug program. uh and if we can now use machine learning to jump uh to to potential solutions uh that could really just really uh kind of push up uh the bar of of what people need to do.

I think it's you take a step back like look at what happened with LLMs. Everyone has to be an AI company now to compete, right?

What's that going to look like in biotech where if you don't have access to a technology like this and then you start to have these certain areas that are closed out to you like what is that going to look like? Um, so I think that that's something that we are uh really excited to figure out.

So if we benchmark it to the LLM market, uh, do you see yourself taking more of an anthropic path or an open AI path? And what I mean by that is uh, or do you develop the drugs?

Do you wind up being the front end to this and kind of being full stack or do you go more like B2B and you have a bunch of pharma companies or you know traditional biotech companies that are buying a product or or service from you? Yeah, it's a great question.

You know, I think if we take any lesson from the LLM market, it's that the winners are going to be the ones who are most agile. The technology is evolving at a faster clip than we ever expected. Our our company goal for this entire year was to get to a 1% success rate on this task.

And you know, we're halfway into the year and we're already past 15%. When you have progress like that, it is really hard to predict like how the market is going to evolve.

And that's why I think just like being flexible, making sure that we have the best technology, making sure that we have the best way to use that technology as well. You know, it's not just throwing a model over the wall and you get a drug.

Uh we really have to make the models work well for the specific applications where they're most important. And we also have to make them much easier to use. We have to get them into the right people's hands.

Uh there's a lot of scientists that if we can bring their creativity and ingenuity together with these new models, that that's where the magic's going to happen. What's the state-of-the-art in like actual the actual impact of antibodies?

Like the last time I heard antibodies in the news was when Joe Rogan was taking monoconal antibodies for COVID infection and uh that was one of those weird things where it felt like a very modern totally not controversial.

Yeah, it felt like a very modern piece of science and the narrative around him was that he was not adopting the modern science but he was kind of on the cutting edge of that fork of the tech tree.

And so is are monoconal antibodies like the most like the latest and greatest or are there other applications that are uh particularly interesting or just driving value even if they're even if they're there are indeed like the traditional way. Yeah.

So about half of drug approvals these days are are actually biologics protein drugs. So antibodies monoconal antibodies are actually just one type. Uh that's closer to what you know we've been exploring in in this paper. But I actually think what's going to happen is that uh people have now started to push beyond that.

They're basic like by specifics are all the rage right now. That's basically you can think of it as two antibodies that are kind of stuck together. So you can actually go after multiple targets at the same time or you can actually get like a more potent behavior for single target.

You know it's kind of just like two things in one. So if we can make you know the fir if we can make each one of those chains like easier to design then you know bringing them together is going to be going to be even easier as well.

So I think it just like raises the bar for the kinds of therapies that you have to work on because it becomes easier to make those drugs. So that's what monoconal antibodies are. They were really important in COVID, right? They can neutralize viruses.

Uh the way that people found them during COVID was they actually went to patients who had COVID. Uh they just like, you know, you would go and just like take someone's blood, for example, and and try to find uh antibodies that are in there and then try to develop that into a therapy.

But now we can just make it on the computer in 24 hours and then go test it in the lab and have some results in in a few weeks. So, there's a lot more work to do to make sure that these molecules will actually work in in in the clinic uh and will work well in patients.

We have early signs uh that that they have drug-like properties, which is really nice to see. Uh but we've got a lot of work ahead of us in order to really bring that to production if if you will.

Uh there's still some more engineering that you might want to do on on the molecules that are coming out of the platform today. Yeah. What what is your sci-fi vision for the state of drug development 50 years from now? And how would the FDA have to adapt regulations to enable that?

Because I can imagine my vision is like you come in to some type of doctor clinic setting, you have some set of conditions and they're just sort of like autogenerating drugs on the fly specifically to your current conditions, but that obviously wouldn't work in in the current um paradigm.

Yeah, I mean that that would be the dream. Um the north star that we have for this space is something uh that we think of as like zero shot drug candidates. Right now we have these zorshock proteins, right? So we specify a target, we get some protein that binds it.

But the fact that this is possible makes us believe that someday we'll be able to actually just like have the models think about all the other properties that we need for a drug. And it's possible that the first sequences that come out of the model will actually be good enough uh to eventually move into patients.

And yeah, if you can do that, then you know, maybe the next stretch is like can we actually do hyperpersonalized medicine, right? Maybe there's something where we take some sample from a patient, we figure out what exact target we need to go after, then we go and generate the drug.

I mean, there's multiple leaps that need to happen. Uh, but I I can't predict what's going to happen 50 years from now. Like, honestly, I failed in predicting what's going to happen in 6 months. I I thought this now you're going to be at 1%.

Do you use AlphaFold or is it just a separate tech that is kind of complimentary but not uh in your direct supply chain? Yeah. Uh, the foundation models are built in house here. So we don't use alpha fold. Uh we actually can't use alpha fold. The latest one is isn't available at the company. So we built our own.

Uh we have a state-of-the-art model uh for predicting protein protein structure and its interactions with molecules. We actually open sourced that.

So we started the company a little over a year ago and then 6 months in that was the first thing we did because you really need like an atomic level microscope uh if you're going to design drugs. That was our thesis when we got started.

um uh but we realized that you know that kind of level of the technology we thought would would get out there anyway. So so we open source that uh it's being used in in most of the major pharmaceuticals today. I think the community has has gotten a lot of engagement around that.

Uh but now the models are becoming even more complicated. So that's why we're starting to invest in not just the model layer but but how we actually make those models useful. Uh they're really starting to look more like pipelines. It's it's kind of like chat GPT 03. It's not just a single inference.

there's actually, you know, a bunch of complex tricks that you need to run if you're going to get the best results. Um, so, uh, that that's that's kind of how we're thinking about it right now. Very cool. Well, congrats on the launch. We'll have to have you back soon. This is fantastic. Yeah.

With a with a rate that you guys are are moving, I'm sure you'll be back on in a couple weeks. So, congrats. Congrats to you and the whole team. We'll talk to you soon. Really appreciate it. Cheers. Thank you, guys. Up next, we have John from Campfire coming on the show. Welcome to the stream, John. How are you doing?

I missed this. You You haven't been hitting the soundboard enough. I love it. Boom. Welcome to the stream. How you doing, John? Can you hear us? Yes, I can. Thank you so much for having me on today. Fantastic. Uh, do you want to kick it off with an introduction on yourself and the company? Absolutely.

I'm the founder CEO at Campfire. We're based in San Francisco and I was a finance exec for about 15 years and just