Google's AI Co-Scientist Solved a Phage Puzzle Humans Missed for Years

Nathan Labenz frames this Cognitive Revolution crossover as a follow-up to his earlier episode on Google's AI Co-Scientist. The tape comes from the Podovirus podcast, where Dr. Jessica Sacher and Dr. Joe Campbell talk with José Penadés and Tiago Costa of Imperial College London. The two labs had recently closed a surprising mechanism. Google's still-unreleased Co-Scientist later put the same hypothesis forward on its own, from literature, without their unpublished data.

Bacteriophages infect bacteria with a tail that docks on a specific cell surface and a head, or capsid, that stores the genome. Phage-inducible chromosomal islands, or PICIs, are DNA sequences that sit quiet in a bacterial genome until a helper phage infects. They then hijack the virus's production line, swap in their own DNA, and burst the cell so that what spreads is PICI, not the original virus — in some cases a collective bacterial defense. The puzzle was a subclass, capsid-forming PICIs, or CF-PICIs, that encode only the head. No tail, and yet they had spread across many bacterial types.

The answer that had eluded the human teams for years, Labenz says, is that these capsids evolved to couple to different virus tails already in the environment. He puts the inference cost of that Co-Scientist run at somewhere between $100 and $1,000. As of the Gemini 2.0 generation, he argues, a language model with scaffolding and a decent compute budget can contribute to frontier science — not only by speeding grunt work, but by supplying an unbiased view or the key insight. Plug in Gemini 2.5, then let systems direct experiments through cloud-lab APIs, and the science-fiction version is already arriving.

Sacher introduces Penadés and Costa as professors who have been working together on mobile genetic elements, especially PICIs. Google let them use a Co-Scientist tool that was not yet public. They gave it the research question and a little background, none of the unpublished data, and watched it come up with the hypothesis they had recently proven. The story had already reached Forbes, The Economist, and the BBC — unusual coverage for phage biology — which is why the hosts wanted scientists, not just headlines, to walk through what the tool is actually good for.

Penadés has worked on PICIs, phage satellites that theoretically use other phages for induction and mobility, for more than 20 or 25 years. For a decade or more the field treated them as defective phages; the lab spent years fighting to establish PICIs as a family of mobile elements. About 10 or 12 years ago they found a peculiar family: capsid-forming PICIs. These carry everything needed to build a small capsid and package DNA into it, but they still need tails to make infectious particles. That is unusual. Other satellites typically hijack everything from the helper phage.

The same CF-PICI DNA turned up in many species — in one example, seven species from five genera. Something was moving these elements between species and even genera. PICI genomes are typically 10 to 15 kilobases, helper phages around 45, so a classic interference trick is to make a small capsid that only the satellite DNA will fit. Costa's lab, coming from conjugation rather than phages, joined to solve the structures. Insertions in the capsid-forming proteins appear to change symmetry and shrink the head so it can harbor only CF-PICI DNA, with a specific terminase that packages satellite DNA rather than phage DNA. They no longer see the relationship as purely parasitic. Some PICIs even carry anti-phage systems; inducing them can cost a helper phage and still benefit the bacterial community.

Imperial's Fleming Initiative put Costa's lab in contact with Google, which was building an LLM system tailored for scientists and wanted researchers to pressure-test it. The first question was not about PICIs. Costa works on conjugation, and a molecular detail of how conjugation actually starts has been unknown for 70 years. They challenged the algorithm; it returned five hypotheses. Checking the top-ranked one, Costa realized, might take many months.

Penadés was in that meeting and saw an opening. The conjugation hypotheses looked impressive enough that he wanted in. Instead of waiting months to validate an AI-driven idea, they flipped Google's intended workflow. They already knew, in the lab, how CF-PICIs spread across bacterial species, and they were close to submitting that story to a journal. Challenge the system with the same question, and see whether it can recapitulate a result that is still locked on their computers.

They were careful about leakage. Because the chimeric particles suggested a toolkit for customizing which bacteria you hit — a way around the narrow host range that limits phage therapy — they had filed for a patent and kept the manuscript off the public internet. Costa says they were reassured the AI could not have seen the unpublished ideas. The top hypothesis was, in his word, astonishing: a recapitulation of what they had observed in the lab.

The mobilization mechanism, Penadés says, is simple once you see it. CF-PICIs produce a capsid with PICI DNA inside. For years the lab's mistake was to assume that whatever is released after phage infection is an infectious particle. What actually comes out, with or without a classic induction, is just the packaged capsid. Free tails are abundant in the environment — phage biologists have known for decades that tails are produced in excess, and purification already showed more tails than infectious particles. The tail that binds a given capsid determines tropism, the host range. Same head, different tail, different bacterium.

Google's top hypothesis was to check whether these capsids can interact with tails from different phages. The system did not explain how that happens, whether it requires induction, or when the meeting occurs. It was the final picture. Many of the new concepts in the human paper are not in the Google document. What *was* there, besides the tails, were the two proteins that decide which tail a capsid can bind: the adapter and the connector. Swap those, and you swap binding ability. The Co-Scientist suggested looking at them. It also suggested checking conjugation as a more promiscuous route than transduction — a hypothesis Costa's lab has started to test.

Costa, whose first love is the adapter-connector interface, notes that some of those proteins bind one specific tail and others bind several, possibly down to one or two residues. That molecular detail is the next story. For the Co-Scientist test, the point is cruder: the system named the parts that actually control mix-and-match, without walking through the cell-lysis-and-extracellular-assembly path the humans had to invent in the lab.

Sacher asks whether the computer's hypotheses were ideas they had already been testing. Timing matters, Costa says: they had been thinking hard about how the machine got the right answer, and it is frustrating, because they had the answer and did not see it. As phage biologists they have known for years that the tail determines tropism. A DNA found in many species might therefore be using different tails. They were biased anyway. After infection, the whole field assumes you have infectious particles — capsid plus tail — for both phages and satellites.

The data had sat there for years. They could induce a Klebsiella CF-PICI, see beautiful capsid-forming replication and packaging, and get no transfer. They blamed the recipient strains for missing receptors, still assuming infectious particles. The same pattern showed up in *E. coli*. A student deleting helper prophages one by one found that removing one abolished even the low transfer while the island was still induced and packaged. That was the crack: maybe one phage induces, and another provides tails. Mixing capsid mutants and tail mutants to reconstitute infectious particles is a 50- or 60-year-old lambda trick. Nobody had thought to take capsids from one phage and tails from another, because the default picture forbade it.

Costa says if Penadés was excited when they understood that adapter and connector were the key, you should have seen him the day the student cracked it. They knew it was big and novel. It took more than a couple of years — "a scary number." Google did not understand the biology in the way a PI does. It made the cheapest connection: you are in many species, to go there you need a tail, so you are binding different tails. Jessica Sacher puts the moral in beginner’s-mind language: a system without a hundred years of phage assumptions can propose what is physically possible. Penadés calls it revenge. He often succeeds, he admits, because he does *not* read too much, and he tells students to trust their controls. On satellites he knew too much. The tails were obvious. He could not see them.

There was another reason to keep the story quiet. CF-PICIs have a broad range, and narrow host range is one of the limits of phage therapy: a therapeutic phage often hits only one species, or some strains within it. Chimeric particles offer a toolkit. Customizing the tail customizes the target, for therapy or for diagnostics. That is why they patented before they talked. Regular lytic phages may do a version of the same mix-and-match; structurally the CF-PICI proteins look like the classic HK97 fold. Anti-phage systems that block tail formation would leave free capsids coming out, and those too might pick up tails and inject DNA into new species. Whether the phage then survives there is a separate question.

When a CF-PICI arrives in a new species, the team has found, it has ways to hijack the phages already there so they are mobilized and transferred as well. Tropism is still set by whether the tip of the tail binds a surface receptor. What adapter and connector add is a neck that is either promiscuous or picky about which tails it will accept. The 2023 paper argued that the capsid genes came from an HK97-like phage and then specialized — portal, terminase, protease all specific — until the connector and adapter, which must bind both the satellite capsid and the helper tail. The practical bet is that you can build synthetic PICIs able to bind multiple tails and deliver DNA across strains or species. That is the patent.

After the first round of five hypotheses, Penadés pushed back. The output was not a real mechanism: it did not say the capsid is released outside the cell, or when the tail binds. Check if this can bind different tails, without saying how, was the level of detail. He told the system the idea did not make sense given how particles are produced. "The reply was crap." The system, he says, does not really understand what it proposed. It connected dots that make a simple story: tropism is tails, so maybe different tails. Everyone in the field already knew that sentence. Readers of the paper will think it was obvious. Somebody still had to say it.

Sacher asks if Google was excited to have a clean test case. Costa thinks the company was pleased to have experimental evidence that the hypotheses were sound and verifiable. Other Co-Scientist-initiated studies in Google's own preprint still have to go through peer review. This lab was further along: they already had the full story, and the system recapitulated it. They joked about stock options. Google did not fund the work. What Costa wishes they had put in the preprint is that for each hypothesis the system listed key papers and notable scientists — about 15 real papers for the tail hypothesis — the trail it claimed to have used.

They also benchmarked the same input against other systems, including Gemini and ChatGPT, as part of the Co-Scientist preprint. The format is different. Co-Scientist returns a comprehensive report with references and names; the others do not. After the experimental paper went on bioRxiv, they challenged various LLMs again. Some could not even find the preprint. One found it and still did not give the right answer. Other systems proposed ideas that were reasonable but not specific — for example, that CF-PICIs take tails already able to inject several species, which would work for any satellite. None said swapping different tails. Some wrote a nice overview of the PICI cycle and clearly had the bibliography. Knowledge is not the same as interpreting something new. So far, Penadés thinks, this is unbiased connecting of dots, not thinking of something completely novel.

Nothing in the experimental manuscript changed. Google's output was already addressed there, and the paper contains new concepts the Co-Scientist never mentioned. What they are writing now is a second manuscript: the input they provided, the output they got, why they think the system reached tails without the bias, and an evaluation of the other hypotheses. That companion piece is preprinted and submitted to the same journal as the experimental paper.

Would they have designed different experiments if they had had the LLM output first? Nothing about the scientific method changes, Costa says. The system is a collaborator. You get hypotheses that are not universal truths; you still go to the lab, run experiments, interpret data, and draw conclusions. What it might change is the path. You might fail 50% of experiments instead of 90%, and save half a year or a year. Penadés's example is all the recipient strains they tested because they thought the defect was in the recipient, when they were simply looking at tailless capsids. They arrived at the right experiments. They failed a lot of the wrong ones first.

Something they highlight in the second preprint: in this example, all five hypotheses were very easy to test. One incorrect idea was that the capsid might bind receptors directly, without a tail; they already had no-tail controls showing the tail is required, and they put that negative in the paper. The system also suggested concrete assays — cryo-EM of adapter and connector, liposomes. Costa is not sure the same ease will hold in other fields. Phage satellite work is mutant complementation and some cryo-EM. A 70-year conjugation mystery is the contrast case: if the ranked hypothesis is even roughly right, it could spare a field an enormous amount of fishing.

Penadés was almost ready to publish capsid-forming PICIs around 2010. Richard Novick, who discovered the first PICI in *Staphylococcus* in 1998 — SaPIs, the founding members of the family — told him not to. For more than a decade people had called PICIs defective phages. Publish a satellite that already carries capsid genes, Novick warned, and the field will just say defective phage again. They waited. They knew the elements existed, that they sat in many species, and that it was odd for a satellite to spend half its genome on genes it could hijack from a phage. Transfer experiments failed. The day a student suggested one phage for induction and another for tails, everything moved fast, because the experiments are simple: mix a capsid with a lysate full of tails, and you get infectious particles, even in natural populations.

The same bias shows up next door. Costa, a conjugation expert, had never thought about conjugation for satellites. They emailed Eduardo Rocha at Institut Pasteur, who has done extensive work on satellites and conjugation, and asked whether he had ever considered it. Never, he said: satellites use phages, plasmids use conjugation. Sacher wonders whether Google trained Co-Scientist specifically to prefer testable hypotheses, an advanced skill even among scientists. Costa does not know. What they do know is the architecture in outline: the system generates hypotheses, ranks them, lets them compete and challenge one another, and scores them with an Elo ranking like chess players. AI experts Costa has heard are surprised by that architecture, especially by indirect hypotheses — CF-PICIs moving by conjugation — that are not sitting on the surface of the available data. Whether that counts as reasoning, he says, is a question for Google.

The system is still under development and not public. Google's strategy, as Costa has heard it, is to fold in more areas of science, test robustness, and eventually release it if it pans out as well as it did in phage biology. The Co-Scientist preprint describes a trusted-tester program; interested labs can contact Google. Imperial will keep the partnership. Postdocs are now using the system in both directions: new projects that need fresh hypotheses, and more established ones. Sacher likes the reverse version — sitting on unpublished answers and asking whether the model can recover them — as a method anyone with a drawer of unwritten papers could run.

A second, unpublished question shows the other face of the tool. Penadés's group also works on plasmids and the impact of transduction on plasmid mobility. Why do plasmids almost never carry *pac* or *cos* sequences, the packaging signals phages use? If a plasmid really wanted to move, it could put one on and "fly." Their answer is that plasmids do not actually want to move too much: a plasmid that transfers too freely may lose variability or burden the host. That less-selfish view is still a small slice of the literature. For years the dominant story has been that mobile elements are selfish.

Ask the system why plasmids lack *pac* and *cos*, and that selfish body of work outweighs cooperativity. The hypotheses come back polished and wrong. They talk about cleavage, about the plasmid becoming non-functional, about damage to the host — the moves a selfish element is supposed to make. The thinking is really good, Penadés says, and it is trapped. They are now in a chat with the system, telling it it is ignoring another kind of relationship with the cell, and waiting to see whether it can re-reason. Campbell, who as a graduate student worked on transposon Tn10 and its multicopy inhibition, recognizes the pattern: even selfish DNA evolves brakes, because jumping without limit trashes the *E. coli* it lives in.

That is the limit that matches the success. When the literature's default was "released particles are infectious," Co-Scientist was free of it and won. When the literature's default is "mobile DNA is selfish," the system inherits the default and loses. Unbiased, in this architecture, means unbiased relative to a human who has lived inside one lab's assumptions — not unbiased relative to the corpus. Penadés's revenge cut both ways. Knowing too much hid the tails. The papers, in bulk, can hide the plasmids.