AI-Designed Viruses: The Future of Antibiotic Resistance?

story published um, in Science this last week and a bunch of articles came out about it. Carl Zimmer wrote about this in the New York Times and there's a great article, if you're not able to get Carl Zimmer's wonderful guest link through the New York Times, um, there is another great article by some scientists in the conversation. Now, what is this story? Researchers published in Science their work titled generative design of bacteriophages with genome language models. What does this mean? Scientists have used large language model technology trained on uh, genomic information about bacteriophages, very specific genetic genetic data sets, and use that to create synthetic bacteriophages. We'll just shorten that and say phages from here on out. Now, it's been simplified and said, "It's a virus." Well, that's because they're like everything that's small and infects things and is, you know, is not really living, quote unquote, it's a virus. It's not bacterial. It These bacteriophages, phages, they're virus viruses. They are or bac- Actually, I'm sorry about that. They are bacteria viruses. Bacteriophages are little tiny viruses that infect bacteria. And they're really cool because we didn't know about them for a long time and then we're like, "Oh my gosh, this is where the predator-prey arms race for antibiotic resistance, all sorts of things." Like we have found our antibiotic targets very often because of how they are um, similar, very important targets that are targeted by these phages very often. And there are even uh, usages in what's called phage therapy for really hard-to- hard-to-kill infections. Where they're using phages. Where they found very specific phages that infect very specific types of bacteria at very specific targets, and they're able to use multiple phages to be able to get rid of multi-drug resistant bacteria. And so this is really important as we're moving into a future without enough really well-working antibiotics. Okay. So, first thing about this story, very interesting use of the what we call a language model, right? They have taken that model system, the training. It's not language. It's genetic code, right? So, they've they've basically done a similar kind of training set. Just the language is base pairs and genetic data. Um and so, they didn't just say, "Here's all the genetic information in the world. Make up a new bacteriophage." They did not do that. They were very, very specific in in their target. What they did is they honed in on one very, very specific natural phage that has a specificity for E. coli. This phage is a Greek letter Greek letter 174. >> [laughter] >> I don't know what I don't know how to say this. Um Omnicom X174. Anyway. Um this phage is a is a template and was the design template for their genome language models that are called Evo 1 and Evo 2. They used this They were all these Evo 1 and Evo 2 models were trained on a bunch of phage genomic data for genetic architecture, specificity to host. So, basically like, oh, this type of structure of information goes with E. coli. This goes with salmonella. This goes with this. Um and so, they created using X174 as a template, they basically created a way to to pump out new designs. Basically, to speed up the process of evolution. So, instead of relying on the natural process of things just bumping into each other and working to infect bacteria and to reproduce themselves using bacterial mechanisms, they don't have to do that anymore. They can potentially make these viral bacteria viruses in a laboratory. They tested nearly 300 chemically synthesized phage genomes. They got 16 that worked. And those 16 were very specific to the host. They were had a had different fitness levels. Um and they had competitive infection kinetics. And so, that's important because it that determines in an ecological system, a bacteria microbial bacterial ecological system, the movement of how different phages are going to be in able to infect or not infect in a certain environment. Anyway, they were able to do this. They uh had generated phages also overcome um bacterial res um bacterial resistance. So, if the bacteria became resistant to phage antimicrobial therapies. They gave them a mixture of different designed phages and the mixture was able to end up with survivors and to be able to outcompete bacterial resistance. And so this is really important for that predator prey interaction that goes on. So, really interesting, great story, lots of cool stuff that is going on with this. The concern is um this was very easily uh managed by uh open source genome language models um and there really aren't any conversations going right now about who regulates and how this kind of techno- technology is regulated if it does become regulated. Where, when, and how is this technology allowed to be used because not only can you use it for therapies and attacking bad bacteria, you could also >> Make a superbug. Yeah. >> You could You could boost up a bacteria. You could destroy good bacteria. You could do so many different things and so there is obviously a concern that comes along with it. >> Mhm. Um I looked it's uh the the first Greek letter is the word is the word phi for fee. >> Phi, thank you. >> And then I don't know if it's X174 or Chi 174. >> what I was like, is this a little >> Yeah. >> [laughter] >> I don't know. I don't know for sure, but >> But hey, there you go. Uh okay, so it's this is all still very proof of concept, it sounds like. >> It is very proof of concept. There are so many reasons this is still far away from actually being used in m- medicine, right? Or in you know, beyond the laboratory at this point in time. This is still you know, years in development. You have to make sure you can test these things. How do we even develop a process for ensuring that these kinds of phage therapies, synthetic phage phage therapies, could be repeatable and um you know, nimble enough and safe enough to use nimbly in a hospital setting or in um in an off-the-shelf you know, how do you make it off-the-shelf? How do you make it something that works really well? Um but it is a I think it's a really exciting um development in the use of AI for science, for medicine, for speeding up the possibility of finding therapies and uses you know, that are you know, otherwise we have to search through all the swamp soup. >> Yeah. Yep. Absolutely. I mean, that's what that's what LLMs are good at is processing data and um creating iterations. Like that's what they're good at because those are the things that we would have to sit and do manually that an LLM can really handle. If you give them a very clear set of parameters they got it. >> Mhm. Yeah, and of course this they said they came up with lots of different options for genetic divert you know, different genomes for phages. 316 of them worked. So, it's not like ooh, it's just going to design a great one every time. There's still testing and iteration and all sorts of stuff that needs to be involved um and the article on the Conversations website does a really good job of going into the gap that is still existent between designing a phage and treating a patient. So, um yeah, but the question is for biosecurity how come we're not how come we don't know about people already having this conversation, right? This Don't Don't Why don't we already have regulations about this? We should. Um anyway. >> Yeah. >> I think that's a a This is my cool news that's going on that everybody should be aware of because this might be where uh bacterial infections are treated in or how they're treated in the future. Um you know, how to and how do we get there faster? It's this is but there are safety issues, so. It's promising. >> Yeah. Yeah.