Anthropic just bought a one-person drug company

In partnership with

When BAIO started in February, we had CEO Dario Amodei’s essays and some early signs of Anthropic’s interest in biology. We still knew little about how far the company intended to go itself.

Eight and a half months later, Anthropic has launched a research workspace, built its own biology lab, bought two biotech companies and begun running drug-development programs. It is working with pharmaceutical and biotech companies on everything from designing molecules to planning clinical trials. This week, we learned that one of its acquisitions was a drug company with a single full-time employee.

To paraphrase the meme: you wake up. Anthropic has done another thing in bio.

In this issue we also have a former OpenAI science chief building automated labs in America, Formation Bio’s $600 million drug-development fund, and early human results for an AI-designed cancer drug. Plus a lively argument about whether AI needs to simulate biology to help cure disease. I have some thoughts on that.

And it’s been a while, but I’ve now added 51 new entries to BAIO’s Tools & Resources Directory, bringing it to 198. They include BioEmu-2, AlphaGenome Atlas and BindCraft2 - plus more datasets, benchmarks and lab tools, with links and access details.

Let’s dive in.

FLYING SOLO

Anthropic bought a one-person drug company

It wasn't the IPO filing, even though Anthropic, a company most people had never heard of two years ago, could be heading for the biggest public offering of all time.

And it wasn't even the ambitious drug-development work Anthropic is now disclosing: helping Bristol Myers Squibb plan clinical trials, Enveda review manufacturing records, and Generate Biomedicines redesign components of cancer-fighting immune cells. The bottlenecks are, as Anthropic pointed out, scattered across the whole process, and the company now has Claude working across them.

No, what really caught our attention the past few days was the one-person drug company.

The one-person drug company that Anthropic has acquired.

We have seen glimpses of this before. In BAIO issues 37 and 55, we told you about Douglas Yao, who used AI, robotics and a garage lab to develop an Alzheimer's drug candidate, then set his sights on building a company with more drugs in human trials than Pfizer.

Now Anthropic is putting its resources behind a drug company with just one full-time employee.

That person is Cody Tranbarger, a biotech investor and operator who spent years posting anonymously as Professor Oak (yes, like the Pokémon professor). Tranbarger has been busy building Mithril Medicine, a company pursuing treatments for two rare forms of muscular dystrophy.

If you’re waiting for an asterisk, here it comes: while Mithril might consist of Tranbarger alone, it is supported by scientific consultants and eleven outside laboratories. Claude helped Tranbarger coordinate experiments, contracts and logistics. The laboratory work still required, well, laboratories, and the scientific decisions still required people who knew what they were doing.

Anthropic bought Mithril in July and disclosed the acquisition this week. It reports encouraging mouse results, with human trials potentially starting in late 2028 if further work holds up.

Tranbarger now wants to “find out how many one person can actually run in parallel,” he writes on X. Joining Anthropic gives his existing programs more resources and him the chance to start several more. He sees an opportunity to spend less of a drug hunter's career coordinating work and more of it deciding which experiments and treatments to pursue.

That also gives Anthropic something to test: how far this way of running a biotech can travel beyond its first operator.

A rare-disease project that struggles to justify an entire company might become more viable when one experienced drug hunter can run it alongside several others. Mithril doesn't establish those economics yet. But it gives Anthropic a working example to build on - and a reason to ask how many potential medicines never get attempted because assembling the company is too much of the job. And how many now will.

BORN IN THE U.S.A.

Former OpenAI science chief launches Warp Sciences to automate drug-discovery labs

The automated-lab lineup keeps growing. The latest addition is Warp Sciences, launched by former OpenAI science chief Kevin Weil with $215 million in funding.

We've recently covered Danaher and NVIDIA connecting AI to laboratory equipment, and Xaira's plans to connect its drug-discovery tools to robots (for a follow-up on that, see the On Our Radar link below). Warp wants to build U.S. labs that drugmakers can hire to make and test molecules.

The location is central to its pitch. Much of this chemistry has moved overseas, Warp says. It wants AI and robotics to help rebuild that capacity in America, with experienced chemists leading the work as automation expands.

“Our ability to lead in science tomorrow depends on the capacity we build today,” the company writes. Its longer-term ambition is “scientific superintelligence,” developed through laboratories that generate evidence for its AI to learn from.

Warp adds a national ambition to the push for automated labs. If AI makes it easier to propose experiments, the capacity to carry them out becomes more valuable. Building that capacity can be both a business serving drugmakers and a way for the U.S. to retain scientific leadership.

Stop typing AI prompts. Start talking.

You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.

Try Wispr Flow Free

THE DEBATE

Does AI need to simulate biology to help cure disease?

What happened: Investment research firm Citrini described a divide between AI optimists and biologists over how quickly better AI could deliver new medicines. Its argument: generating molecules can be fast; getting the experimental evidence to judge them is much slower.

That prompted a provocation: The X user Mira argued that enough computing power could, in principle, simulate biology exactly and find solutions without experiments. Validation would still be needed, but “experiments are a way to save compute,” the post claimed.

The pushback: The post drew a wave of objections about whether computing power could substitute for missing biological evidence. Among those pushing back was Morgan Levine, formerly VP of computation at Altos Labs, who asked where the information for that simulation would come from. “Infinite compute can’t help you simulate a system that’s never even been observed or measured,” she wrote. Her further objection: many evolutionary histories could fit the same constraints. Why would a simulation recover the particular biology that actually evolved?

But…: Personally, I don’t care in the slightest about simulating biology exactly. I do care about curing disease. Those are different goals. An AI could help us find treatments by choosing better experiments and learning from their results without producing anything close to a complete simulation of life.

The commenters are right to challenge the claim Mira actually made. But disproving that claim wouldn’t settle how much AI could accelerate the search for treatments.

And about the missing data: I’m less convinced by the implied pessimism about getting it. AI could (and already does, to some extent) help decide which experiments to run, robots could make more of them affordable, and better models could extract more from the results. That won’t make every kind of evidence quick or easy to obtain. But today’s data shortage doesn’t, by itself, tell us how long tomorrow’s will last.

A bitter lesson? In AI, the bitter lesson is that methods which learn and search at scale have repeatedly overtaken systems built around human expertise. I suspect there are some in store for biologists. The possibility I have in mind is that systems learning from experiments will find useful patterns and interventions that our narrow explanations of biology would never have led us to.

Which actually brings us back to Morgan Levine: She has now written a whole essay about how AI might learn to perceive biological structure that our human categories obscure. Instead of teaching machines to organize biology as we do, she argues, we could let them learn from biological observations and discover patterns our familiar categories miss.

Sound familiar? In May, we reported from Levine’s Vitalist Bay talk. She argued for collecting diverse views of living systems, including signals humans cannot yet interpret, so AI could learn the underlying biology.

MORE SHOTS ON GOAL

Formation Bio launches $600 million fund for clinical-stage drugs

When we covered Formation Bio in issue 18, its argument was that drug development had become the bottleneck: finding more candidates would achieve little without the capacity to test them in people.

Now the company has announced a fund with $600 million in commitments to help acquire and develop 10–15 drug programs over the next 3–5 years.

Formation buys or licenses drugs in early clinical development and uses AI alongside its drug-development team to move them forward. Each funded program will sit in a separate company, sharing Formation's scientists, clinical teams and infrastructure. The aim is to put more money into trials without rebuilding a development organization for every drug.

The fund's backers include research foundations, endowments, hospitals and philanthropic organizations. Formation is investing alongside them.

Formation is putting financial weight behind the thesis we covered in April. If AI produces more candidates, funding and running the trials could become an even tighter constraint. This fund gives Formation more opportunities to test whether its development model can turn that growing supply into medicines.

THE AI WAR ON CANCER

Insilico reports early tumor responses to AI-designed cancer drug

Insilico Medicine has reported tumor shrinkage in patients taking its AI-designed drug for mesothelioma, a cancer of the tissue lining organs such as the lungs.

The pill, ISM6331, was designed using the company's Chemistry42 platform and targets TEAD proteins, which help control gene activity. In the ongoing Phase 1 trial, four of 22 evaluable mesothelioma patients responded across the reported dose range. All had previously received treatment.

Insilico highlighted a 40% response rate at the highest reported dose, but that means two of just five evaluable patients. Three of the four responses across doses were still ongoing at the August data cutoff, including one lasting about a year.

These are company-reported results from a small study without a control group. Insilico is still selecting a dose before a planned Phase 2a study.

Founder and CEO Alex Zhavoronkov also hinted at a wider ambition on X: “this molecule was designed to work in a broad spectrum of biological conditions beyond cancer and TEAD has a role in aging.”

In issue 39, we covered Insilico's lung-fibrosis drug rentosertib entering Phase III. This is a different molecule, in cancer, at a much earlier stage. It adds another human test of the company's AI-designed chemistry.

☑️ Molence releases Mol-1, an open model for predicting biomolecular structures.

☑️ Xaira demonstrates a robotic workflow for protein production and analysis.

☑️ UK commits £100 million to an AI biosecurity lab, with support from OpenAI and Anthropic.

☑️ Pythia Labs reports AI-designed antibodies that activate or block hard-to-target receptors.

☑️ Pathos raises $237 million in a Bristol Myers Squibb-led round at a roughly $3 billion valuation.

☑️ Boehringer Ingelheim and Microsoft explore digital patient twins for clinical-trial design.

☑️ CellBridge uses AI to identify cell types in pathology images.

☑️ Hugging Face and collaborators release Carbon-X to answer questions about DNA sequences.