Anthropic's AI lab claim splits biologists as protein discovery debate widens

Anthropic's molecular biology lab, staffed by Claude agents, flagged a repeating pattern surrounding a known enzyme. The company says the pattern had not been catalogued before. The claim, first announced last week and dissected in MIT Technology Review on 28 September, has turned into the sharpest public argument yet about what AI can legitimately claim to have discovered in the life sciences.

MIT Technology Review described the system as 950 agents working for 21 hours. What they produced was not a new DNA sequence, the outlet reported. They found a repeating pattern around a known enzyme, and Anthropic's announcement described it as "reminiscent" of the work that led to CRISPR. That word choice did a lot of work, and it is where the trouble started.

The bar problem

The critique arrived quickly, and from inside the field. A viral post from biologist Lucas Harrington, later endorsed by the chair and CEO of Eli Lilly, argued that "finding a weird cluster of genes and repeats is often the easy part. The hard part, and where the real discoveries come from, is figuring out what the system actually does." MIT Technology Review quoted the post directly. Harrington's own suggestion, as the outlet reported it, was that AI companies should "set the bar high now, so that when an AI actually discovers a fundamentally new biological mechanism, everyone appreciates how big a deal it is."

Then the story got messier. Mario Rodríguez Mestre, a biologist at the University of Copenhagen, said over the weekend that his team had already found the same pattern, according to the New York Times, as relayed by MIT Technology Review. Mestre had regularly used Claude in his work, and he asked whether Anthropic's team had learned from his conversations. Anthropic denies that. Mestre says he has stopped using Claude anyway.

"The hard part, and where the real discoveries come from, is figuring out what the system actually does."

None of this means the underlying work was worthless. MIT Technology Review makes the point that narrowing 200,000 candidates down to a handful worth testing is genuine scientific labour, and that a general purpose chatbot managing it is notable even with humans steering and running the experiments. The problem is the frame. Once the question becomes whether Claude itself made a discovery, there are only two answers available: breakthrough or bust. Everything in between gets flattened.

What protein work actually looks like right now

The Anthropic fight matters beyond one company because protein and materials discovery is where AI has been quietly accumulating real, checkable results. The pattern across the dossier is less dramatic than a press release and more useful.

Recent headlines logged alongside the Anthropic coverage include work on thermostable mRNA lipid nanoparticle vaccines using what Nature described as data-efficient AI, published on 28 September. Separately, researchers reported using AI structure prediction to speed the discovery of molecular glues, small molecules that can force proteins together in ways that treat disease. Neither of those claims rests on an AI being credited with a conceptual leap. They rest on AI doing a large amount of searching and ranking faster than people can.

That distinction is the whole argument. AI systems are excellent at the part of science that is essentially triage, and the field has been slow to say so plainly. MIT Technology Review noted the same dynamic in mathematics: earlier in September, OpenAI said its agents had cracked a million-dollar problem, and within a couple of weeks the discussion had shifted to whether that particular problem was the one mathematicians cared about. The outlet was careful to note the criticism did not claim OpenAI's solution was wrong, only that the result may not have been the important one. There was also an accusation from a mathematician that the models used some of his work without credit.

So the dispute is not really about correctness. It is about significance, and about who gets to decide. That is a harder thing to litigate in public than a wrong answer.

Verification is arriving from a different direction

While the discovery-credit argument runs, other groups are working on the more mundane problem of whether AI-generated biology can be trusted at all. Headlines from the last week describe DeepMind watermarking proteins so labs can verify what they are looking at, and separate work on SynthID Bio watermarks for AI-designed proteins. These are attempts to make AI output traceable rather than to make it famous.

The commercial side is moving at the same unglamorous pace. Hansa Biopharma said it is partnering with Cradle on AI-driven protein design and engineering. This is contract work, not a breakthrough claim, and it is probably a better guide to where the technology actually sits than any announcement about agents discovering something.

Anthropic's own framing is part of the problem, according to MIT Technology Review. The outlet argued that AI companies are not presenting their systems as tools in the way microscopes or supercomputers are tools. They are insisting the systems make discoveries themselves, which sits awkwardly with how science usually produces new knowledge: through collaboration and an expanding set of instruments. That mismatch does not just annoy biologists. It makes genuine progress harder to recognise when it does arrive, because the only available verdicts are the two extremes.

The rest of the science file

Elsewhere in the dossier, the week's research news is unusually concrete. Scientists at the Harrington Discovery Institute at University Hospitals and Case Western Reserve University published findings in Cell Reports describing a molecular mechanism that may explain why lost weight returns. The work centres on asprosin, a hormone released by fat tissue that signals the brain to stimulate appetite. In mice, obesity altered fat cell activity in a way that kept asprosin elevated after weight loss, an effect the researchers call obesity memory. A separate laboratory led by Seth J. Field confirmed the cellular change persisted, according to SciTechDaily. The authors caution that the human evidence comes from publicly available datasets and still needs direct confirmation.

Astronomy produced a cleaner first. A paper submitted to arXiv on 15 September reports the first direct detection of auroral radio emission from an exoplanet, the giant planet beta Pictoris b, using the MeerKAT array. The authors, Kevin N. Ortiz Ceballos, Edo Berger and Yvette Cendes, report rapid, recurring and highly circularly polarized bursts at frequencies of 0.85 to 3.5 GHz. They identify the emission as electron cyclotron maser radiation, which implies a magnetic field of at least 1.25 kG at the planet. The abstract states plainly that no radio detection had previously been unambiguously localised to an extrasolar planet rather than its host star.

That is what a discovery claim looks like when the measurement does the arguing. There is a signal, a physical mechanism that explains it, and a number that follows from the mechanism. Nobody has to be talked into it.

Why the Anthropic fight will not settle soon

The structural reason is that AI companies are in a race with each other. MIT Technology Review closed its analysis by noting that as Sam Altman and Dario Amodei compete, raising the bar for what counts as an AI scientific breakthrough may be the last thing on their minds. Lower bars are better marketing. Harrington's proposal, that companies hold the line high so a real result lands properly, asks them to act against their own incentives.

There is also a practical cost to the noise. When a pattern-finding result is billed as a CRISPR-scale moment, the biologists who might otherwise engage with the tool have to spend their time explaining why it is not. Mestre's response, abandoning Claude over an unresolved question about whether his conversations fed the system, is a small data point about how quickly trust erodes.

For now the honest summary is narrow. AI systems are producing useful, verifiable results in protein and materials work, and the dossier contains several examples of that. They are also being credited with discoveries in ways that working scientists reject. Both things are true at once, and the gap between them is where the next argument will happen.