What did Google's AI Co-Scientist "Discover"? The Human Scientists' POV, from the Podovirus podcast

Hello, and welcome back to The Cognitive Revolution. Today, we're following up on our recent episode on Google's AI Co-Scientist with a special crossover episode from the Podivirus podcast, in which hosts Dr. Jessica Satcher and Dr. Joe Campbell speak with José Pinares and Tiago Costa, the scientists at Imperial College London who recently made a surprising discovery, which Google's AI Co-Scientist later put forward as a hypothesis entirely on its own. For context, here's a quick crash course on the biology that you'll hear discussed in this episode. Bacteriophages are viruses that infect bacteria. In general terms, these phages reproduce by inserting their genes into a host bacteria cell and hijacking the cell's protein-making mechanisms to produce copies of the virus itself, until the cell ultimately bursts and releases many copies of the virus into its environment. Structurally, phages consist of a tail, which is specifically adapted to attach to specific types of bacteria cells, and a head, also known as a capsid, which stores and protects the genetic material until it's injected into a target cell. Phage-inducible chromosomal islands, also known as PICIs, are really a fascinating product of evolution. They are DNA sequences that have evolved to lie dormant in bacterial genomes until the cell is infected by a certain type of bacteriophage, at which point they become active and actually hijack the virus's reproduction process, replacing the virus's normal DNA with copies of itself. The affected cell still ends up bursting, but instead of releasing copies of the virus that infected it, it releases viruses that spread the piggy DNA to its sister cells, in some cases thus serving as a collective bacterial defense against the attacking virus. Now, the question that Jose and Tiago and their teams, and also independently the Google AI co-scientists, had set out to answer was how a certain class of PICIs, known as capsid-forming PICIs, or CfPICIs for short, which encode only the capsid or head portion of the virus, with no tail to latch onto other target cells, had somehow managed to spread widely across many different types of bacteria. The surprising answer, which had eluded the human scientists for years, but which Google's AI co-scientist was able to surmise just from its analysis of the relevant literature— at an estimated inference cost, I should say, of maybe somewhere between $100 and $1,000— is that these capsids have evolved the ability to connect up with different kinds of virus tails in the environment, and that's how they were able to infect and ultimately become incorporated into many different kinds of bacterial cells. Beyond serving as a vivid reminder that evolution is an eternal arms race and that we should absolutely avoid evolutionary competition with AIs at just about all costs, this episode shows that as of the Gemini 2.0 generation, large language models with proper scaffolding and a decent inference budget can now contribute to frontier scientific research. And not just by expediting the grunt work, but in some cases by providing an unbiased perspective or even The key insight. Obviously, such hypothesis generation can and will accelerate scientific discovery, even if its hit rate is ultimately fairly modest. And you can imagine how quickly this becomes even more powerful as Google plugs Gemini 2.5 into the co-scientist architecture, which they've surely already done by now, and then again as AIs begin to direct experiments and collect their own data via Cloud Lab APIs. This is all science fiction, But it's happening for real in our lifetimes right now. And once again, I can only conclude that the singularity really is quite near. As always, if you're finding value in the show, we'd appreciate it if you take a moment, share it with friends, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Your feedback is always welcome too. Feel free to reach out anytime via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. Finally, before diving in, I should note that Eric Torenberg, in his new role at a16z, has just announced that he's actively hiring podcast hosts across multiple domains, including biology and biotech. If you'd like to have conversations like this for a living, definitely check his Twitter for more information on the opportunity. With that, I hope you enjoy this foray into microbiology and this early glimpse of AI-powered scientific discovery. With Dr. Jessica Satcher, Dr. Joe Campbell, and Professors José Pinares and Tiago Costa from the Podivirus Podcast.

Hello, welcome everyone to Podivirus Podcast. So today we have a special episode, and I might say that every time, but we're doing lots of different things all the time these days. And I heard about this exciting AI story that's also a phage picky story. So it was the obviously perfect choice of topic, and I'm so glad to have convinced José Pineda and Tiago Costa, who are professors at Imperial College London. They have been working together, I'm sure we'll hear a lot more, but they have been working in this space of mobile genetic elements, phages, specifically phage-inducible chromosomal islands, or PICIs. And they were working on this and Apparently Google let them use its not yet released Co-Scientist tool, which is an AI tool. And they were able to use it to kind of, sounds like, rederive a body of work that they had not yet published, but had been working on for a couple of years. And giving the AI tool this research question and a little bit of background, but none of their unpublished data, they saw that the scientist was able to kind of come up with the same hypothesis that they had recently proven, but not yet published. So this is a story that we want to dig into, and you might have seen a bunch of coverage of this lately. It's been on Forbes, The Economist, BBC, so they made it to the mainstream, which doesn't happen all the time with phages. So I'm sure, yeah, you'll be able to look into that coverage too, but we wanted to kind of talk to our scientists, our phage scientists especially, but everyone who might be curious about what are these AI tools useful for and how do they actually start using them? So I'm very excited to have you both here. And of course, my co-host Joe Campbell is here too. So we'll get deep in and hopefully find out more. What are the pros and cons of using this kind of tool and what are the limits right now? To start, tell us just a little bit about what were you studying when you came across this tool? How did that even come into your lap? Okay.

So for, I don't know, for more than 20, 25 years, we have been working with these pKs that they are phage satellites. So they are using theoretically other phages for induction, for mobility. And it was, so yeah, for at the beginning we have quite a few issues because people just thought that these pKs are defective phages. Okay. For Probably for more than 10 years, we were fighting just to put the pikis as a kind of new family of mobile genetic elements, as fate satellites. And probably, I don't know, 10 years ago or 12 years ago, we discovered a peculiar piki family that there were these kind of capsid-forming pikis, or what we call CF pikis. And they were peculiar family because they have all the genes required to produce the capsid, the small capsids, and to package the DNA into the capsids. But they still require TALEs for creating infective particles. This is quite unusual because the majority of all the satellites, the other satellites just eject probably everything from the phage. Okay. But this is a very peculiar thing. So we realized that these capsid-forming PKIs were present exactly the same DNA.

Okay.

The same element in many different species. Sometimes we put as an example one capsid-forming virus that was in 7 species from 5 different genera. So it was a bit surprising how these elements— we thought that probably this is a new mechanism of gene transfer allowing these elements to go between species and even between genera. And this is the things that we have been doing in the lab for a few years to try to understand the mechanism.

Exciting. And then were you also working with Tiago's lab, or how does that come into play?

So Tiago is a structural biologist. Probably he will, he can introduce himself later. It's quite funny because he works in another important mechanism of gene transfer. He's an expert on conjugation. We want, we really wanted to know how these capsid-forming piques, these small capsids that have the ability to package the pique DNA inside how they look like in terms of structural, you know, because theoretically the genes involved in the production of the small capsids, the capsids forming piquis, the capsids are related to phage proteins. But somehow they have evolved the ability to produce small capsids and just to package the piqui DNA inside. And so we start working with Thiago because he's the expert in getting all these kind of structures and things and In fact, he solved the mystery, but I think it's better he introduce himself to explain.

Go ahead.

Yeah.

Well, we solved a fraction of the mystery, so we still don't know the full story, but I think we are in the right track also. And this is also supported by the AI co-scientists that they seem to agree with our experimental data and our hypothesis. Yeah. So, so I come from a different, different mechanism of gene transfer from conjugation. But I mean, my lab and Jose's lab share this, you know, this love and excitement for how bacteria either directly transfer the DNA among themselves by conjugation or how they use the phages to mobilize the DNA. So Jose, well, asked me to, or we have teamed up to understand from different angles the mechanism of how this capsid-forming are able to widespread in different bacterial species. So my role, or the role of my lab, was to understand at the molecular level how those, all those CFpKs are formed structurally, how these capsids are able to, you know, take one protein, a phage protein, and change it to make these structurally very different entities. So just to go a little bit more in detail, we understood that there are some insertions of protein sequences into the capsid-forming proteins that dictate, we think that dictate the different symmetry of these capsid-forming units and make them smaller so that they can only accommodate the DNA of the cfpt and not the the DNA of the phage, which sort of makes it smaller so that they can only— are able to harbor the DNA of the piki.

Probably it's better to— we didn't say, you know, usually the satellite, these kind of pikis, they have 1/3 of the size of the phage genome. So usually these pikis are around 10 to 15 kilobases and the phage, the helper ones, are around 45. So a classical mechanism of phage interference is for the satellites to produce small capsids. where just the pQ or the satellite DNA can be packaged.

Got it.

So yeah, and this capsid-forming, they all just produce small capsids. They also have specific terminases to package the capsid, the pQ DNA, not the phage DNA. So this is kind of very specific packaging system to, you know, for to package the capsid-forming DNA into small capsids that are formed by proteins encoded by the satellite. That was another unique thing compared to other satellites. Yeah.

And I think I saw in one of your papers that these are not considered parasites because they're not having a detrimental effect on the phage. Is that right?

So for many years, we— I think it probably was in terms of selling the papers, it was nice to talk the war between the parasites, you know, and— but in this case, the prototypical member of this family we didn't see an interference with the phage reproduction. So the phages, you know, we don't know for other capsid formings if that's the case or not. But even we published, I think was 2 years ago, even some phages that might suffer for inducing the PKIs. Because, for example, we published that because some of the PKIs have anti-phage systems. So even if you have some cost inducing these PKIs, if you can move them against other phages, you can also benefit. So, we don't see now them as a parasitic relationship, might be, but we might see more, we see them more as a, yeah, it can be more synergistic or depending on the scenario. They are more friends than we thought probably 15 or 20 years ago.

Okay? Okay. Okay. So, what were you What made you want to use this co-scientist AI tool? Where were you when you figured out that was an option, and why did you set it to work on this problem?

Thiago started using it for his work. Thiago, go ahead.

Yeah, yeah. So yes, so actually there's a Fleming initiative at the college that put us in contact with Google, and because they knew that Google was developing this new AI LLM model tailored for scientists, and they wanted to engage with the scientists so that they could understand how good the LLM model was. And so actually initially this was not related to Piggy. So there was a, the work started with my lab trying to understand particular question in terms of conjugation that has been unknown for, for 70 years. So we were interested to understand how conjugation is actually initiated. Okay. And we wanted to understand the molecular determinants behind the T= 0 of conjugation. So we have challenged the algorithm And our team came up with several hypotheses, 5 different hypotheses. And we realized that would take us many months to understand whether the hypotheses, the highest ranked hypotheses would be correct. Okay. So, so José was on that meeting and immediately saw an opportunity. There. Okay. So why we don't challenge the algorithm with the actual, our experimental data that we already have in our hands and we were about to submit, or I don't know if we were about to submit or already submitted to, to one of the top-tier journals.

Okay.

So, so, so, so, so basically we turned upside down the strategy of Google instead of having an AI-driven hypothesis that would be validated in a few months down the line. We knew already in the case of the PKIs, we already knew the mechanism behind the spreading of the CFPKs in different bacterial species. And we have challenged the algorithm with that exact same question. And it was astonishing. I mean, one of the top or the top hypothesis was basically a recapitulation of what we have observed in the lab.

Wow.

So were you surprised? Yeah, go ahead, José.

He didn't say, but the hypothesis that the scientists suggested for the conjugation— I'm not an expert there, so I'm just— he was really impressed. So that's why I said, okay, this looks really good. I need to take some advantage here because if the hypothesis provided by the co-scientists for conjugation were Probably I wouldn't have been interested, but he was really positive. Okay, this is looking very good. So we did this kind of, we changed the approach and it was quite impressive. Probably the audience probably would like to know now that, so the way this capsid forming can be mobilized between different species is very simple.

Okay.

They can produce these capsids with the piki DNA inside. Okay? And that was a mistake we did for many years. We always thought that whatever is released after phage infection or whatever, it's an infective particle.

Yeah.

What we realized in this paper is that these pikis can be induced or even without induction, as soon as the cell dies, what is released is just the capsid. picky with the package DNA. And outside the cell, they have the ability to take tails from different phages, free tails.

What?

I mean free tails. Okay? And then depending on the tail, that will determine, because we know for many years that the tails in the phage biology determines the tropism. Okay? So that's it. Okay. So there were many things on the paper, okay, until we realized that—

Yeah.

It's induction or non-induction. The cells lyse, okay. The capsid, the packaged DNA will be released, no tails, and then outside will be free tails. We knew for many years as well that there were many tails that produce in excess. Okay. And even during the purification, they already showed that there were more tails than infective particles or whatever. And then it's for the same capsid, depending on the tail, that will bind there that will determine the tropism. Okay? So, and then one of the hypotheses that Google says is like, you need to check the possibility that these caps can interact with different tails from different phases. Okay? The system didn't talk about how that happened or that was in the outset. So, it was just the final picture. Okay? So, many things that was in the middle, It's not— so many new concepts that we put in the paper are not on the Google document. But the final thing, that was right. Okay? And it was something also quite impressive is that we knew for the genetics, and now Thiago is solving the structure, hopefully will be soon, that there were 2 key players in this, the ability of a capsid forming and binding to one tail or another, is 2 proteins called the adaptor and the connector.

Okay.

So, if you swap the adapters and the connectors, you also swap the ability of a capsid forming to bind to some tails or others. Okay? Core scientists also suggested we should look at the adapter and the connector.

Wow.

Okay. So, it was so many things that were not, you know, in the output from the core scientists. But, the key one, the tails, that was there. Okay.

Wow.

And it was another one. So we put this, you know, that was quite interesting is they suggested, the system also suggested to check conjugation. That was quite funny, you know, as a new mechanism because conjugation is more promiscuous than transduction theoretically because of the tail, the narrow for a specific tail. Okay. We have been, we have started working on that as well to see. if that's true or not, but that was quite an interesting hypothesis as well, okay, that this system provides.

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

Thiago must have been excited that it chose your favorite area.

Yeah, well, it goes back to the beginning, right? That was the first we have tried was to understand a bit better the—

I sense a question. Thiago loves the adapter and the connector.

Ah, yeah.

You have dinner or lunch with, you just need to mention the adapter because it's pretty cool that some adapters and connectors bind to one specific tail, you know, look to be very narrow and others have the ability to bind to several tails. Okay. From different species. And he's all the time very excited to see how mechanistically Probably one or two residues, you know, he's— that's the words you need to mention to Thiago. Connector and adapter. That's it.

I love it.

I'm just going to get a little better understanding of the process. So, I assume because you hadn't done this when you started the experiments, were the hypotheses that were generated by the computer things that you had thought about and therefore you designed experiments to test them? Or did you have ideas? I mean, how were the experiments designed by you that ended up leading to the conclusions that agreed with the AI-generated ideas?

I agree with that because the timing is very important here.

Yeah.

So, yeah. So We have been thinking a lot, okay, how this machine got the right answer.

Okay?

Right. And it's very frustrating. It's very frustrating because we have the answer there and we didn't see it. Okay? As a phase biologist, we know that the tail determines the tropism. Is that right? Yeah, we know that many years. Okay?

Yeah.

For years. So, If one DNA using a tail is in many different species, it's because maybe it's used different tails. But we were biased. That's the problem we had. And I'm killing myself for that because for many years, I always thought, and all phage biology people will think, that after infection, what you have are infective particles.

Yeah.

With the capsid and the tail. And we knew it's very things, very strange. We have PKIs that can be induced, but we didn't get transfer. We couldn't understand this thing because we were wrong in the way we solved this thing. We were so biased. Okay? We always thought that after induction or whatever, it's released from a bacterium are infective particles, both for phages and both for satellites. Okay. And I think that's the big thing. So Google didn't understand what's happening here, but make the most simple connection. You are in many different species and to go to these species, you need a tail. You are binding to different tails. That's it.

So if I could just go back to the order, am I hearing you say that you had data and you weren't exactly sure how to interpret it, and then you realized that this was telling you how to interpret it?

No, we had the data for years.

Yeah.

Okay. So for example, we have the PK, we use the capsid form from Klebsiella.

Okay.

And we induced the PK. So this is a helper prophage. Until the paper will be published, for us, a helper prophage means a phage that will provide whatever is required to produce an infective particle, capsid and tails or just tails. Okay? So we have a strain in Klebsiella 1 we induce. We show very nice capsid-forming replication, very nice packaging, no transfer.

Hmm.

So in our mind was because the recipients we use didn't have the receptor. But we were thinking all the time that there were Infective particles. Okay. And the same happened to us in E. coli. And then one day, suddenly, talking to my student, we realized that in E. coli, there were several prophages. Not— but when we induced them, we knew that the island was heavily induced with a lot of replication and a lot of packaging because we can purify the capsids from the lysate. But the transfer was very low.

Okay.

And then, he was making mutants to delete, it was, I think, 6 prophages, one by one, to see what's the inducer, whatever. And then, he realized that it was one day that he removed one, eliminated the transfer, even the low transfer, but the island was still induced. And, at some point, we thought that maybe it will be phages to induce and phages to provide the cells.

Okay.

And then we start connecting the dots. And it has been known for 50 years or 60 years that you can get lambda-80 lysate, mutagenic capsid, mutagenic tails. You induce them, you take the lysate, you mix, and then you have infective particles.

Hmm.

But the people that did these experiments, they never thought taking capsid from one phage and tails from another. Never. Because we all thought that everything, you know, that's the big bias. Okay. So then that day everything started making sense.

So if you think that I was excited when we understood that adapter and the connector are the key factors here, you should have seen José on that day when he cracked this hard nut. He was very excited because he knew that this was something big. Something novel.

And you did this before the AI came into play. You cracked this, but you were spending a couple years trying to get there.

And then—

More than a couple.

Okay. More than a couple.

Yes.

Many years, a scary number. And so then the fact that the AI came up with that because it did not have this bias that once you release a particle, that of course it has its tail, it has whatever it needs to be infectious. It was unbiased in that thinking. And of course, why not? Could you have this later meeting of Cap'n and Tail outside the cell? And that's sort of what happened.

So, so, yes, sorry, Jose. I think we need to also to clarify a bit the chronology here because there was another angle that was in play here, which obviously we, we saw opportunity to patent. these new elements because these elements are broad range, okay, which is one of the limitations on phage therapy. You have all these narrow range, these phages used in the phage therapy, they either only hit a bacterial species or sometimes strains within that species. And with these chimeric particles, pretty much you tailor-make and open up The, you know, you have a toolkit where you can customize the target bacteria that you want to hit by for therapy or to make diagnostic tools. So we kept this very secretive because unless the— we could only speak about this in public once the patent was filed.

Right.

So, so yeah, so there was no public domain. So that's why we are very reassured that AI system will would never have access to this manuscript or these ideas because they were kept in a safe spot in our computers.

Yeah.

Also, we challenged other systems as well with the same input. Okay. And no, none of the other systems, you know, even closely related even to the scientists, you know, and even, so I think the important thing, we didn't mention this thing in the preprint. When we got the first round with the 5 hypotheses, we asked, because as I said, it's not really a real mechanism in terms of the capsids will be released outside the cells and they will, you know, was just, okay, check if this can bind to different cells without telling you how this interaction will happen or when will happen or whatever. Okay.

Yeah.

Like, make a second round. And I said, yeah, but this is, you know, challenging the system saying, yeah, but this doesn't make any sense because how these particles are produced, where will be whatever. And the reply was crap. You know what I mean? So, the system don't really understand what it's proposing.

Yeah.

It's just connecting the dots, which makes sense. You know, just taking this simple thing that the tails determine the tropism. Maybe it's because it's using different cells. That's it. And it was there. We all knew that. Okay? So, this is the type of things that for the people, when they read the paper, they will think. Every single person will understand very quickly because it's obvious. And, you know, but somebody needs to tell you. You know what I mean?

Yeah.

So, it was quite frustrating because it was quite, you know, a few years trying to understand. We have strains and a lot of mutants, and we couldn't understand why we didn't get transfer. It's because just this tailless— we said this is a new entity that we call the tailless capsid forming, you know, the packaged DNA. These are the entities that are released. Okay.

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

Can phages, regular lytic phages, have this sort of mix and match of tail and capsid, or have you started to look? That's so cool.

If I did a bit, I will say yes.

Yeah. So they all have this—

We thought that was the same, you know?

Yeah. They have these adapter connectors. Like, do many phages have those? Is that conserved?

Yeah.

Interesting.

We have asked—

Okay.

Sorry. I think the biggest difference is that when a phage infects, the capsid and the compatible tail are already there. So we think that the number of capsids that will be released probably Lower. Okay. But structurally, the capsid forming, the proteins involved are exactly the same that you have in the typical HK97. And this is the things that Tiago— but also, we think that even other myoviruses or whatever, they work in kind of similar thing. The funny thing that we are working now, and probably the people that— it's good that we will say them publicly because then if somebody copy us, can always get the credit. So, you know now that there are many anti-tail, for example, anti-phage systems that block the tail formation. So when that happens, you will have capsids.

Oh, it's freely coming out.

So we really think the same will happen with— okay. And the phages will have the ability to inject DNA in different species by swapping tails. If these phages in the new species survive or not, this is another question. Okay? We also show that the capsid-forming, when they go to another species, they have mechanisms to hijack other phages that are in the new species to also be mobilized and to be transferred. But mechanistically, we think that will be the case. Okay?

Yeah. So tropism is you're like narrowing, kind of getting more high resolution on what it's actually defined by, like not necessarily just the tail, but Specifically, this connector point maybe a little bit.

Yeah, I mean, the true true pit means will be always determined by the the deep of the tail, right? Whether it binds the surface or the receptor in the target cell. But it's true that now the neck region where the portal, the adapter, and the connectors are will have a same. in determining that tropism is not by the direct binding to the surface or the receptor, but by generating a promiscuous or a very specific structure that binds only one or several phage tags.

That's so cool.

And that's what we are looking after, what we sort of start to understand now. What is imprinted in those structures of these proteins that determine that specificity or that promiscuity in to bind only one type of tails or a set of tails?

It's—

but this is for the next story. This is the next story.

Yeah.

Yeah.

This capsid formid, so we propose in the paper that we published in 2023, That became— the genes came from DHK97 fate, let's say, you know? But they have evolved. So this is not crosstalk between the 2 proteins, even they are very similar in terms of sequence. Okay? So you have the capsid, the portal, the terminases, you know, the proteases are very specific. Okay? So everything is very specific until they arrive to a point that is the connector and adapter, that each connector and adapter needs to bind To the capsid, the shear peaky, or the phage peaky. And then they need to bind to the same tail. So it will be a quite funny thing. You evolve something to be completely separate in terms of capsid-forming things and the proteins involved that you cannot crosstalk or you cannot, but then the tail is the same. And probably it will be a competition between them as well. So it's quite interesting to know. what's happening there, you know? Because the idea we have is that probably we can create synthetic Pkis with the ability to bind to multiple tails. And if that's possible, we will limit, you know, we will solve one of the limitations for phase therapy, you know?

Yeah.

We can create, we can deliver DNA to multiple strains or multiple species, or, you know, this is the pattern that Tiago was mentioning before.

Yeah.

Yeah, I think this is so cool to just— it feels like one of the first fresh sort of phage-host range discoveries that I can remember for a while. It's an actual another level of looking at it as a whole. And pairing it with the AI aspect, it feels like we have this potential tool, set of tools that we're all going to, it's going to be more, not less. Like we're all going to probably start using these tools. And now we can have someone almost like someone in a beginner's mind, like who doesn't have all the assumptions of our 100 years of history. And then now we can get their input into just physically what would be probable or possible. And now we can kind of bounce off that whole other set of knowledge, different. Differently built than ours.

I agree. For me, something that is kind of revenge, you know, no one nice for me. So, I normally, I don't say these things because people said that this is not the right things to say, but I think I succeeded. Some of the things we found in the lab was because I don't read too much.

Yeah.

Okay? So, I'm not biased. Okay?

I love that.

Yep. Yeah. Yeah. To some extent.

You can't admit it. It's hard to admit.

I admit, you know, and so many times, you know, I have a very good idea in Galápagos. Wow, this looks so good. And then I go to Medusa. Yeah. Somebody in 1985 thought the same. And then you have a huge admiration for these people, you know, that, but I wasn't biased, you know, that happened, for example, with the lateral and these kinds of things, you know, because I was—

Interesting.

are completely ignorant. I always say to my students, I trust your results, because the students sometimes said, José, but— I don't mind what people said. You have the controls, it works well, let's interpret your results. Okay? And now I was so biased. I knew too much about the satellite and I had this thing there, the tails. It was obvious. You know what I mean? So, I think that's the good advantage for this kind of systems, that there's no bias. The system, it doesn't matter if this is a phase or it's a conjugation. We always thought that the satellite should move by a phase because that's the name. We never thought, you know what I mean? But why not by conjugation? It's just to put a very small sequence. So it was a good revenge in that way that the system gave me in that way because I knew too much and I was so biased, or we were so biased when we talked about this thing. And then, yeah, we realized. And another funny thing is that I have no idea about tails, okay? When, yeah, zero. It's something that I never ever thought about, that the tails will be something interesting at all. That was a tail, it's formed. The funny things were the packaging, the capsid. This is— the big papers from the capsid, ISCA97, you know, the terminate, you know, and now suddenly, it looks like the tail. It's a funny thing as well, you know. It's because, as you said, you can have now a very simple mechanism to move things between species, you know.

Wow. Yeah. So, was Google kind of excited that you guys showed the world a nice test case? Like, it's pretty Neat how well it worked. Is that, is it really that clean cut? They must have been very excited.

Yeah. Well, that's a question for them, obviously, but I think they were very pleased to have experimental evidences that, that hypotheses that the system was generating were sound and solid and verifiable in the lab. I mean, there, there was other studies that, that they, they were initiated by hypotheses. From the co-scientists that are also reported in their system preprint, but they still have to go through the peer review and then publication at the end. In our case, we were a little bit way more advanced because we had already a full story. We knew the answer and the system was able to recapitulate.

We tried to claim some options from them, but, you know, so it's like as a joke. Because—

what was that?

Some stock options from Google.

Oh yeah. I was going to say they should be funding your lab.

They didn't, they didn't agree with us.

Okay.

Because I think that they were lucky and we were lucky as well, you know, because I think this is the perfect system to test because as I said, everything was there. Okay. You can produce the capsid, you can produce the tails, whatever people have done experiments in the past, mixing things, you know, tail mutants with capsid mutants, and then you have an infective particle. So. All was there, just putting the dots with OpenBio. So I think it was lucky for them and lucky for us, you know?

Yeah, definitely a very nice partnership between the—

Yeah.

Well, what's that? Oh, there was no funding involved, so they didn't invest a penny because—

Exactly. We are getting money from Google. We don't. We are trying, but so far—

So, so, so there was no, no, yeah. So, so it was mutually benefit for both sides.

I don't know if you checked our preprint because we didn't put, sorry, Joe, we didn't put in the preprint, but even the system for each of the hypotheses provides key papers that the system used to arrive to that hypothesis. You think it would be nice, for example, to, to include that information as well for the people to Because they were mentioning names, as you know. And for each name, they highlight one paper that they thought was relevant for the hypothesis. And that was pretty cool as well. You know, so it's like for each hypothesis they run, there were like, I think it's a few papers.

Yeah.

It was for the tail things, for the hypotheses we were mentioning, I think there were 15 papers that they say, okay, this is the papers they found relevant to arrive to that hypothesis. Okay.

That were real papers.

Yeah, real papers. A lot of sense, you know?

Yeah.

Can you tell if the other AI tools that you used that weren't as effective, did they see those papers and not realize the importance of them, or did they just somehow miss them? Or I guess in trying to understand what's different and why they're giving different answers, I guess I could. I broadly think of 2 categories. One is that somehow they didn't find the right papers, or they found them and they didn't realize, weren't able to understand the importance. So, can you go back and figure out if those other search engines even, or whatever learning tools or AI tools even looked at those papers?

Yeah. So, so the— so part of the preprint, we had to benchmark the system, the CoI system, with other systems. Okay, so Gemini, ChatGPT, and the output from those systems are different. Okay, so the format is not the same. However, the hypotheses and, and the answers are there, but some— the output— well, the Co Scientist output outputs the references, he outputs notable scientists that have worked on that field. So this is a very comprehensive report that you obtain. However, other systems, they do not give such a comprehensive output because they work differently. CoSciENtist is very— I mean, we are not AI people and it's for— still CoSciENtist, at least for me, it's still a black box. But the system is built differently. And funny enough, we have published the preprint of the experimental paper in bioRxiv. And then we challenged the system, the different LLMs. And some, they haven't seen the preprint.

Right.

They could not even find our already published data. Well, it was— What?

No, I think what the system found the preprint, but didn't provide the right answer.

Okay. So it's the same. It didn't provide the right answer even though it found it.

But they found the preprint. I think that was one case. And in other systems, they also propose things that make sense. So for example, some of the systems proposed that maybe This capsid forming can take tails that have the ability to inject DNA in different species, the same tail. Okay? But for example, in that case, that will work for both, for this capsid forming or for any satellite. You know? So it's not something that is specific because at the end, the tails are very similar. So some of the hypotheses were kind of okay, but none of the other systems provide the right thing, you know, swapping differentials. And, you know, and even some of the systems, I can't remember which one was the last one we put in the paper, they made a very nice overview of the PKS cycle, you know. So, it was definitely had very good access to the bibliography, but I don't know to what extent, as I said before, that's a problem. Because this is the problem we have. You have a very good knowledge, then how you interpret things that You know, it's something new. And that's the question that some people ask to us, you know, to what extent these are just new hypotheses or just linking, you know, connecting the dots. And I think so far it's more connecting the dots in an unbiased way, way more than thinking something that is completely novel. Okay.

Yeah.

If I could ask just a couple more questions. First is, are you in any way planning to Rewrite your manuscript that you hadn't shown it based on what it did. So are you, I guess my guess would be it would most likely affect the discussion, but are you planning to rewrite it? And well, I'll ask, let you add to that before I ask the follow-up.

What? So for the experimental manuscript, nothing changed. Okay. So basically what the output from Google, we already addressed in the experimental manuscript. Even in the experimental manuscript, we provide new concepts that are not in the— co-scientists never mentioned. What we are doing now is like, because we thought that it would be nice to share our experience, we are creating a new manuscript that is our experience with the co-scientists.

Right.

Explaining what's the input we provide, okay? And then evaluating what's the output that we got, okay?

Yeah.

And explaining, for example, the tails, this is a good thing, but this is things where explaining this kind of frustration or why we think EcoScientist was able to arrive to that point, even this non-bias, even did they provide the picture? And we also make kind of evaluation from the rest of the hypothesis. We preprinted that manuscript as well. And we sent, we have sent the manuscript to the journal as well, to the same journal that did the experimental one, because they might be interested in having both the experimental one and this coscientist-related one.

I guess the other thing I was thinking about, and maybe this is a hard one, is If you had done this, had the output of this LLM system before you started the experiments, I guess the hard question, what would you have done differently, if anything? Would you have done different experiments or no?

Okay.

Nothing changes.

Nothing changes. It's just a paint. Yes. So the way that I see the system is it's like a collaborator, okay? It's something that you can interact with. You obtain hypotheses in the end, but they are not the final truth. You have to go to the lab, run the experiments, interpret the data, and draw conclusions. The hypotheses that the system generates are not universal truth. You and the scientific method does not change at all.

Right.

What it does, it puts you perhaps in the right path right from the beginning. So if you— Right.

I understand that the hypothesis doesn't affect that you need to do an experiment to address it, but I guess what I'm asking is, I mean, Back when I used to do experiments, often if you have a hypothesis, you design an experiment to test it. And were there things— so when you said it was quicker, is that because you would've designed the right experiment?

Maybe, yeah, because you have not— maybe you would've not failed 90% of the experiments that you have failed, but you just fail 50%, and this would save you half of a year or 1 year of experimental work and 1 year—

So it would've changed things in terms of it would've changed which experiments you did and didn't do, right?

Exactly.

It would change—

I'll give you an example. Coming back to the idea that we knew that we have some strains that we can induce, the ILN, it's been induced, it's highly packaged. And we tried to get transfer and we didn't get any transfer because there were just capsid, no tails there. So can you imagine how many recipient strains did we try? Because we thought it was a defect in the recipient strain. Okay? Capsid with these reads, these— can you imagine how many? And we didn't get an answer for that. And then we moved to E. coli, the same thing. And then start deleting these things, deleting, because we never thought that these were tailless.

Okay.

Or we never thought that, yes, if they have everything to package the DNA, just need a tail, maybe they can bind to different. So, it's like, at the end, we arrived to the right experiments, but we failed a lot of the experiments because we didn't understand. what was happening.

Right.

Yeah. If they give you some kind of path and the others, there are other hypotheses that you realize that might be relevant for the biology of the satellites, but they were not really important for the question because it can apply to many other satellites. So there was no exclusive of this family. But the type of conjugation is the same. They provide a hypothesis, That is very easy to test. Something that we highlight in this second preprint is that at least in our example, all the 5 hypotheses are very easy to test.

Yeah.

And also, as Tiago said, we had the feeling that we were talking to an expert in the field. Even, you know, says, what do you think about this? I think you should check this thing. Okay, that makes sense. Definitely. At the end, we were very happy because we arrived to the conclusion, the same one. But for the other examples, for example, the one that Thiago was saying, it's a mystery for 70 years and nobody thought about it. Can you imagine how many experiments people did in the past trying to understand the time zero for conjugation? This is a big question. I can't believe that these guys working on conjugation don't really know that. Okay. So the system provides you a hypothesis that looks okay. It's saving a lot of time.

Right.

Yeah.

I wonder if in your paper about this method, you could sort of give concrete examples of of why you would get moved from 90% of the experiments not working to maybe only 50% of the experiments not working.

Power.

And sort of say, we would've done, if we had known this before, we wouldn't have gone down this rabbit hole and we wouldn't have gone down this rabbit hole. And because, I mean, I guess maybe I'm trying to think about if someone's out there trying to say, well, why am I going to do this? I mean, I think a lot of people would be attracted to know which experiments that you did, which ended up being dead ends, you think you might have just not have done if you had run this AI prior to starting the experimental work. Does that make any sense?

Yeah, it makes sense.

Yeah, it makes sense. But this, yeah, I mean, you tend to for— to forget the failed experiments and keep track of the good ones, right? So yeah, yeah, I don't know.

You—

and that's why we arguably— you don't, you don't publish the negative results, or maybe you should, but anyway, that's another question. But yeah, it's— I think it's now, it's a, it's very difficult to, to unders— to say which experiments—

Yeah.

Have been done or should have not been done or could have been done differently.

Yeah, you know, I'd be smart, but I mean, I guess it's a question of whether that—

But if you're with the AI, this won't be an issue because if it will be an AI-driven hypothesis, you have never run those experiments before. Okay. So in this case, you, because we knew the answer, we had experimental, we had the portfolio of experimenters of failed and successful. We could do that judgment. But in the future, if it's an AI-driven hypothesis, all the experiments, I think that's what will make a good or a quick or quicker discovery is how good the human will be to critically interpret the hypothesis and design an experimental setup Although the AI system already gives you some experiments that you should, that you could do to test that hypothesis. And that will, the human will have definitely a very important role still in this process.

So, I think I can answer your question. Okay? So, we knew these capsids for, I don't know, for how many years.

10 years.

almost ready to publish this typeset for Minpikis in 2010. And Richard Novick told me, José, you can't. Richard Novick was the person that discovered the Pikis in a stamp that he calls Happis. This is the first member of the family. Let's say he's the father of the Piki. He discovered the first element. And as I said, for many years, he discovered, I think it was 1998, and for more than 10 years, 12 years, people thought were defective fetuses. So when we discovered this capsid forming, that was probably 2010. And he said to me, he said, you can't publish this thing because then people will think that these are defective fetuses. They have everything, they have the capsid, they have whatever. So you can't. Okay. So we wait. So it has been for many years. We knew they exist. They kind of manual, you know, findings. We knew that they were in many different species. We couldn't understand why satellites want to use half of the genome to carry genes that they can hijack from the phage. So it was kind of things. We did a lot of experiments that did not transfer or whatever. And the day talking to the students that we realized that maybe some satellites use a phage for induction and another phage for the tail, everything made sense.

Okay.

And I think it's when Thiago was mentioning before, I said, Thiago, I think we have something pretty big. And since then, everything was very quick, very quick, because these experiments are very simple. If you have a capsid mutant, you can have the tails, you can mix. So these things even work in natural populations. So a normal phage, prophage, when it's induced, there is a lot of tails there that if you just mix with the capsid forming, you have infecting particles. Everything We have an experiment in the paper. We have a donor for the capsid for miR-8, a donor for the prophase, a functional one, and a recipient. And things move, you know? So, but that was the day that we realized that maybe this is something for induction and something to provide the tails. And maybe we realized that, you know, everything makes sense. So you have something from Google that said, maybe this can use different tails. you're thinking, because the experiments are very easy to do, at least in our case. I don't know in other areas. So I think it's because of the bias. And I said, it was awful because we didn't see the big picture. But having something that is new, it's like the same thing in conjugation. We never thought, I never thought about conjugation.

Okay.

Tiago never thought about conjugation. He's an expert on conjugation. There is a guy in Pasteur Institute, Eduardo Rocha, He's an expert on satellites and conjugation. He has made a lot of programs and studies about ORETs that are required for— and we sent an email saying, Eduardo, have you ever thought about this? And he said, no, never. Because satellites use phases, plasmids use conjugation. You know, this, you know, so we are so biased that things that are there, you know. we didn't realize. So, I think that's the main, I think, the power of this system, at least for us. It's no bias. Whatever has been published, they will highlight that. And then, it's up to, as Tiago said, it's up to you to decide if that makes sense or not. And, for the hypotheses they provide, were very easy to test, all of them.

I want to touch on that last point. I think the testability is really interesting. And I wonder if behind the scenes, the AI co-scientist that Google made, maybe they gave it a lot of training related to knowing what science is like when it, so it could understand like what would be a testable hypothesis versus not. Because I don't think that's usually, that's probably not, you know, what you're getting when you talk to regular ChatGPT or other LLMs or even regular scientists. Like, I think that's like an advanced skill to think, you know, of all the things you could do, what are the things that are the most bang for your buck and the most testable, least time in the lab? Like you have to have a lot of background to know that, but even filtering on that like lens and giving you the hypotheses and giving you testable ones, like it's pretty interesting that you said they're also testable. And I wonder if that's built into the system. And if so, that seems like a really nice.

Yeah.

feature. Is that your view of it?

The thing is, I don't know. I think this is a question for the Google team. The way that— I mean, we do know the basic principles. We know that the system generates different hypotheses. They are ranked and then they compete and they challenge each other and they are ranked with a Elo ranking, like the chess players. We know that it's a very complex system because I've I've heard experts on AI talking about the co-scientists, and they seem to be very surprised on the architecture of the system. And in particular, these novel hypotheses that have— that Jose was just mentioning about the ORETs and that these CFETs could also go be mobilized by conjugation, this what they call it indirect assumptions or hypotheses and that they are not imprinted directly or they cannot readily interpret by the data that is available. So whether there is an element of reasoning in the system that would come up with these less evident hypotheses, yeah, maybe. But this is again something for the Google people.

This system sometimes provides a specific You know, for example, for the connector and the adapter, they said you should check by cryo-EM how these things look. For some vesicles as well, the system suggests to use liposomes. So it was quite a funny thing. Even some of the hypotheses they found, even though they were incorrect, they were quite funny because we included them in the paper as a negative. So for example, one of the hypotheses was that just the capsid forming, there were proteins there that might have the ability to bind to some specific receptors without the tail. Okay. And we, you know, so it's like a different mechanism of entry. You say, well, this is a— and we include these controls because we have experiments with no tails to show that the tail was absolutely required. Okay. So—

Okay. So you ruled it out.

Yeah. It was some kind of suggesting experiments, you know, in a very easy way. Okay. But I, again, I have the feeling here that the fate wall is a very easy It's a very good model for the system because it's just, at least the work we do in the satellite is mutant complementation disease from cryo-EM. Yeah. We don't really know within other areas, you know?

Yeah.

Wow. Well, yeah, I think we should probably let you go on with your evening and close out, but this has been so, so cool. I guess my last thing I wanted to end on was just, when can others use this, or are you still going to be using it as a still kind of in a testers-only phase, or what's that like going forward?

So the system is— it's been just under development, right? The system is not publicly available still. This is, again, that's the grounds of Google. But the discussions that we have had with them is that there is a process for the maturation of the system till it gets publicly released. Okay. I think the strategy will be to incorporate different areas of science to the system and understand how robust the algorithm is. And eventually, if it pans out as well as at least with phage biology, it will eventually be released to the public. I think now it's still, this is written in Google's Co-Scientist preprint, is that there is a test, trust-tested program, and the labs interested in testing the system can can contact Google.

Yeah.

We will definitely keep working with them because we have a very close collaboration and partnership with them. And I believe it will be mutually beneficial for Google and for us as it has been so far.

We are using it now with some postdocs, okay, that they are in the process, very talented postdocs. So to challenge the system with the questions they really want to address.

Yeah.

At different stages. So there are projects that are quite new, so to see how this system answers the questions and how the system's better established. So yeah, we are having some kind of, but this is a different approach. It's now asking for new hypotheses, you know, and we will give some time.

Forward version.

Yeah, it's a forward version.

I like your reverse version. I think, yeah, both make sense, but it's like forward and reverse genetics, like the reverse AI co-scientist, like It makes so much sense. Everybody else is sitting on data that hasn't put it out there.

We have done another thing that it's quite funny. You know, I don't know. So we work also with plasmids and the impact of transduction and plasmid mobility. Okay. So we have a paper now, again, nobody knows about the paper. So why you gotta have PAX or COS sequences on plasmids? Because if as a plasmid, you really want to move, just put a PAX or COS. And then that's it, you will fly.

Okay.

So we have an answer for that. It's like in reality, plasmids don't really want to move too much. Okay. Because then we'll— if there is a plasmid that moves too much, it will lose variability or whatever. Okay. But this idea of the plasmids being less selfish than we thought, still not— it's a very small proportion in the literature. So far, for many years, we thought that the mobile are selfish. Okay. So when you ask the system why there is no Cas or Pac sequences on plasmids, that body of literature is more important than the cooperativity or whatever. So most of the hypotheses that are pretty good are wrong because that's just being selfish. You know, they do these things, the terminase will cleave the plasmid and the plasmid will not be functional. Or, you know what's more, on the thinking that was a selfish element. Okay. It was something that blocked the biology of the plasmid as planned. So it's a quite funny thing, you know, the thinking is really good.

Yeah.

But yeah, because most of the literature is based on the selfish thing. You can see that from the hypothesis. Okay. So the system also needs to learn probably what papers are more important, you know, the recent one versus the I don't know. So this is, we are working there, you know, in some, we try against the reverse for the plasmid thing. The hypotheses were pretty cool, but they were wrong. And we are working now.

Did you try the other LLM systems with that one?

No, for this one. You can do that. Is that what we announced?

I mean, I guess it could be scary, but could be interesting if, depending on the question you ask, different LLMs work better.

That's a good point. And we are now having the chat and we are telling the system, yeah, that's really cool, but you are ignoring that these are not really selfish. You can have some kind of another type of relationship with the cells. So let me know what you think if you include that possibility as well. And we are waiting, you know, so this is, it's quite interesting. Yeah.

I guess you wanna be selfish, but within reason, because if you're dependent on Bacteria for survival. If you get too selfish, you might kill your own host. Like, I don't—

Yeah.

But when I was a graduate student, I was working in a lab that studied this transposon TN10, and it had this thing called multicopy inhibition because it, you would naively think that it just wants to keep jumping and making more and more, but eventually that's bad for the cell. Right. So it has a mechanism that prevents it them transpose, makes it transpose less when they get too many in the cell. And it was a really— so that's like, I was thinking about that when you were talking about it.

Yes.

Because again, that's another selfish DNA that you think is like, the best thing would be for it to jump whenever it can. But if you start jumping too much, you trash the E. coli that it's in. And then, Correct.

Yeah.

All right.

AI is going to find Joe's thesis soon and—

Well, that's not what— that was someone else's work.

Well, thank you. Thank you all so much and good luck with everything. And I can't wait to see your papers come out and we'll be posting them and yeah, can't wait to get this out. So keep going.

Thanks a lot.

Yeah. Thank you. Thank you for the invitation. Okay, bye-bye now.

So welcome. Bye.

Bye.

It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.