What did Google's AI Co-Scientist "Discover"? The Human Scientists' POV, from the Podovirus podcast
Summary
- Google’s AI Co-Scientist independently surfaced the central hypothesis that Imperial College scientists had spent years proving. Given the research question and public literature—but none of the team’s unpublished data—it proposed that capsid-forming PICIs spread by attaching their packaged DNA capsids to different free phage tails outside the cell. The opening narration estimates just $100–$1,000 of inference cost, making this a striking low-cost demonstration of hypothesis generation.
- The system’s edge was freedom from a field-wide assumption, not complete biological understanding. Phage researchers assumed material released after infection was already a complete infectious particle; Co-Scientist instead made the simple connection that “if the tails determine the tropism, maybe it’s because it’s using different tails.” Yet it never reconstructed the release mechanism, and when the researchers asked for that detail, the system’s reply was reported as, “Crap.”
- The experimentally supported mechanism could become a programmable delivery platform with broader reach than conventional phages. CF-PICIs release tailless capsids, then use adaptor and connector proteins to accept tails whose tips determine bacterial targets; swapping those proteins changes tail compatibility. Tiago Costa says this creates a potential toolkit for tailor-made therapeutics or diagnostics across strains and species, addressing phage therapy’s typically narrow host range, though the full structural story remains unresolved.
- The credible productivity claim is fewer dead-end experiments, not automated science. Costa first said the experiments themselves would not change, then clarified that an early correct hypothesis might cut failed experiments from “ninety percent” to “fifty percent,” potentially saving half a year or a year. “The scientific method does not change at all”: humans must still assess hypotheses, design experiments, interpret results and reject attractive errors.
- Co-Scientist’s scaffolding materially outperformed the general-purpose LLMs tested on this problem. Gemini, ChatGPT and other systems produced plausible summaries or adjacent ideas, but none identified the same different-tail mechanism—even when one apparently found the team’s experimental preprint. Co-Scientist returned five ranked, readily testable hypotheses, notable researchers and roughly 15 relevant papers for the winning tail hypothesis, creating the feeling of “talking to an expert in the field.”
- The result is a clean validation case, but not proof that the system routinely discovers novel mechanisms. Penadés describes its current strength as “connecting the dots in an unbiased way” rather than inventing something wholly new; the dots were unusually complete, including decades-old experiments mixing separate capsid and tail mutants. A second reverse test on plasmid mobility produced sophisticated but wrong hypotheses because the literature’s dominant “selfish element” framing outweighed the team’s newer, less-selfish interpretation.
- Near-term access and investability remain gated by Google’s development process. The system was still unavailable publicly, with outside labs directed toward a trusted-tester program while Google tested robustness across scientific fields; no funding or stock options went to the Imperial team. The Cognitive Revolution’s narrator frames this as a Gemini 2.0-generation capability and speculates that plugging Gemini 2.5 into the architecture, followed eventually by Cloud Lab APIs, could extend such systems from proposing experiments to directing them, but that remains the narrator’s forward-looking scenario.
Deep dive
1. CF-PICIs posed a host-range mystery for more than a decade
Penadés has studied phage-inducible chromosomal islands for 20–25 years, including a long fight to establish PICIs as phage satellites rather than defective phages. Around 2010, his group encountered capsid-forming PICIs: elements encoding their own capsids and DNA-packaging machinery but still requiring phage tails to become infectious.
The size difference is functional. A typical PICI genome is roughly 10–15 kilobases versus about 45 kilobases for its helper phage, so CF-PICIs build smaller capsids that accommodate their own DNA while excluding the larger phage genome; Costa’s structural work implicates protein-sequence insertions that alter capsid symmetry.
The puzzle was distribution: Penadés cites one exactly identical CF-PICI found across seven species from five genera. If phage tails normally impose narrow tropism, how could one element traverse such distant bacterial hosts?
The relationship is not necessarily parasitic. The prototypical CF-PICI did not visibly impair helper-phage reproduction, while some PICIs carry anti-phage systems that can impose a cost on phages when the PICIs are induced. Penadés now sees them as potentially synergistic—“more friends than we thought probably 15 or 20 years ago.”
2. A 70-year conjugation question led to a reverse validation test
Imperial College’s Fleming Initiative connected Costa’s lab with Google while it was developing an LLM system tailored to scientists. Costa initially asked about the molecular “T equals zero” that initiates conjugation, a question he says has remained unresolved for 70 years; Co-Scientist returned five hypotheses whose leading candidates impressed him, although testing them would take months.
Penadés saw an opportunity to invert Google’s intended workflow. Rather than start with an AI hypothesis and await validation, the team asked a question whose answer they had already established experimentally: how CF-PICIs spread across bacterial species.
Their safeguard was unusually strong. The mechanism was absent from the public domain because the manuscript and patent-related ideas had been kept “in a safe box in our computers,” so the system could not simply retrieve their answer. Costa emphasizes that Co-Scientist recapitulated an existing human discovery; it did not generate the unpublished experiments or originate the completed story.
Penadés calls the case mutually lucky: Google received experimental evidence, while the researchers obtained an independent test of their reasoning. The collaboration involved no Google funding—despite jokes about requesting stock options.
3. The winning hypothesis was extracellular tail swapping
Penadés’s mechanism starts with a conceptual change: when cells lyse—after induction or even without induction—the relevant entity is not a finished infectious particle but a tailless capsid containing packaged PICI DNA. Outside the cell, that capsid can capture free tails produced in excess by different phages.
The attached tail then determines where the particle can deliver its DNA. One capsid can therefore acquire different tropisms depending on which compatible tail it encounters, explaining how the same element reaches multiple species and genera.
Co-Scientist put this possibility at or near the top of its five hypotheses: “You need to check the possibility that these capsids can interact with different tails from different phages.” It also independently directed attention to the adaptor and connector proteins that genetics had identified as decisive; swapping those proteins swaps which tails a capsid can bind.
The system nevertheless supplied only “the final picture.” It did not explain whether the components met inside or outside the cell, how release occurred or several other concepts developed in the experimental paper.
4. Human researchers were trapped by what every phage biologist “knew”
Penadés’s candid diagnosis is bias: “We knew too much and we were so biased.” Researchers knew that tails determine tropism, yet assumed everything released following phage infection already combined capsid and tail; consequently, failed transfer looked like a receptor problem in recipient bacteria rather than evidence that the donor produced only capsids.
In Klebsiella, the team observed strong CF-PICI replication and packaging but no transfer, then tried many recipient strains. In E. coli, a student deleted roughly six prophages one at a time; removing one abolished even the low transfer rate while PICI induction continued, separating the phage responsible for induction from the phage supplying the tail.
The component logic had been available for 50–60 years. Researchers could induce lambda or ϕ80 capsid and tail mutants, mix their lysates and recover infectious particles—but, as Jessica Sacher notes, they had not made the leap to capsids from one element accepting tails from another.
Once Penadés and his student proposed separate induction and tail donors, “everything started making sense” and progress became fast. A simple experiment combining a CF-PICI donor, a functional prophage donor and a recipient demonstrated movement, including under population-like conditions.
5. The discovery’s long gestation sharpened the lesson about bias
Penadés says he was nearly ready to publish CF-PICIs in 2010, but Richard Nobig—the discoverer of the first staphylococcal PICI, which he called SaPI—warned that capsid genes would reinforce the old misconception that PICIs were defective phages. The team waited while accumulating years of replication, packaging and non-transfer results.
Penadés normally values ignorance as protection from received wisdom: “I don’t read too much. Okay? So I’m not biased.” He tells students to trust well-controlled results over what the literature says, but here his deep satellite expertise produced exactly the blind spot he usually avoids.
Sacher likens the AI contribution to a “beginner’s mind” unburdened by 100 years of history and assumptions. Penadés agrees, while narrowing the claim: Co-Scientist did not truly understand the system; it connected two familiar facts—wide distribution and tail-determined tropism—without respecting disciplinary boundaries.
6. Adaptors and connectors could turn host range into an engineering variable
Costa distinguishes true cellular tropism from component compatibility. The tail tip still binds the bacterial receptor, but the portal-adaptor-connector neck determines whether a capsid accepts one particular tail or a promiscuous set, indirectly controlling which tropisms become available.
Penadés says the team proposed in 2023 that the CF-PICI genes came from HK97 phage machinery. The capsid, portal, terminase and protease components have evolved to be specific, while the connector and adaptor must bind both the capsid-forming PICI and the relevant tail. The full mechanism remains unresolved.
The researchers speculate that perhaps “one or two residues” could determine narrow versus broad tail compatibility; Costa’s lab is beginning to investigate what those structures encode.
The patent opportunity is a synthetic PICI able to bind multiple tails and deliver DNA across multiple strains or species. Costa frames this as a customizable therapeutic or diagnostic toolkit that might mitigate phage therapy’s narrow range, not as an already demonstrated product.
Penadés suggests that ordinary lytic phages may also swap tails, particularly when anti-tail defense systems block tail formation and leave capsids. He keeps the caveat explicit: delivery into another species may occur, but whether the incoming phage survives there is “another question.”
7. Co-Scientist’s architecture produced better questions, not final answers
Benchmarking against Gemini, ChatGPT and other systems yielded different formats and plausible alternatives, such as a tail with the ability to inject DNA into different species. None reached the key different-tail mechanism; one system apparently found the experimental preprint yet still failed to extract the answer, while another demonstrated excellent command of the PICI lifecycle without making the decisive inference.
Co-Scientist’s output was more comprehensive: ranked hypotheses, relevant scientists and supporting literature. Penadés recalls about 15 real, sensible papers behind the tail hypothesis, giving users an inspectable route into the evidence rather than an unsupported answer.
Costa understands the system only at a high level: candidate hypotheses compete, challenge one another and receive an Elo-like ranking. He says AI specialists appeared impressed by the architecture, especially its capacity for less direct inferences such as possible CF-PICI mobilization through conjugation.
Testability was a practical strength. It suggested cryo-EM for adaptor and connector structure, liposome experiments for a vesicle idea and a wrong hypothesis that capsid proteins might bind receptors without tails; the team’s no-tail controls ruled that last proposal out. Even incorrect candidates could become crisp experiments.
8. The scientific method survives, while literature bias remains the failure mode
Costa’s operating model is “like a collaborator”: hypotheses are not “universal truth,” and researchers must still experiment, interpret and conclude. Responding to Joe Campbell’s pushback, he clarified that early guidance could change experiment selection—perhaps reducing failures from 90% to 50% and saving half a year or a year—without changing evidentiary standards.
Penadés makes the savings concrete: the correct framing could have prevented repeated recipient-strain screens for particles that lacked tails. Yet reconstructing the precise counterfactual is difficult because laboratories remember successful experiments better than failed ones and rarely publish negative results.
A second reverse test exposed the system’s limits. Asked why plasmids lack phage pac or cos packaging sequences, Co-Scientist generated thoughtful but wrong selfish-element explanations; the team’s newer interpretation is that plasmids may avoid excessive movement and preserve diversity, a less-selfish framing underrepresented in the literature. Campbell’s Tn10 analogy—multicopy inhibition prevents transposition from overwhelming its E. coli host—supports the broader intuition.
Co-Scientist remained under development and unavailable to the public, with interested laboratories pointed toward Google’s trusted-tester program. Penadés and Costa are now using it prospectively with postdocs at different project stages, testing whether its “forward version” can retain the reverse experiment’s value when nobody yet knows the answer.