One prompt, four follow-ups, 64 hours of unattended work. Anthropic's research model came back with 211 protein designs for the lab, and 17 of them grabbed scorpion toxins. What it didn't do is make an antivenom.
What Anthropic asked Claude to design
Scorpion venom is a mix of small proteins that jam the sodium channels running through nerve cells. The standard treatment, antivenom, comes from immunized horses. It takes months to produce, covers a narrow set of species, and needs a cold chain that most rural clinics don't have.
Anthropic's target was narrower than a cure. The team wanted small binding proteins, called minibinders, that attach to specific toxins and block the spot where they grip a sodium channel.
Five toxins from four regions made the list: Cn2 from Mexico, Ts1 from Brazil, AaH I and AaH II from North Africa, and LqhIII from the Middle East.
The work ran inside Claude Science on an internal research model. The project ran with Citrini Research, a financial research firm, as a partner. The technical report and all 211 designs are public on Hugging Face under Anthropic/scorpion-toxin-binder-design.
Downloadable designs matter here. Most AI drug discovery announcements arrive as a press release and a bar chart.
Inside the 64-hour autonomous campaign
People supplied one opening prompt and four follow-ups. The model handled the rest: which tools to call, which batches to fold, what to plot, when to write up results. Humans watched. They didn't drive.
Across the run it used 12 design tools and produced 318,484 protein designs. It folded and filtered 553,052 design-toxin pairs and burned about 10,000 GPU hours doing it.
Then came the part that separates this from a generative demo. The model chose 180 designs to send for wet-lab testing. The human team added 31 of its own, which brought the shipment to 211.
A model picking which designs deserve lab money is a small decision with a large consequence. Wet-lab slots are the bottleneck in protein design, and something has to rank the queue before anyone spends money at the bench.
How the 17 binding designs were verified
Two contract labs ran the assays separately. Adaptyv Bio used a cell-free system. Twist Bioscience expressed the proteins in human kidney cells as Fc fusions. Both measured binding with surface plasmon resonance, which reports whether two molecules stick and how tightly.
A design counted as a hit only if it cleared four bars: binding in both labs, signal in every replicate, a dissociation constant at or below 1 micromolar, and no binding to CLEC12A, an unrelated protein used as a control.
The hits landed unevenly across the target list.
Toxin | Region | Designs that bound |
|---|---|---|
Cn2 | Mexico | 7 |
AaH II | North Africa | 4 |
AaH I | North Africa | 4 |
LqhIII | Middle East | 2 |
Ts1 | Brazil | 0 |
The two best designs, Cn2-B1 and Cn2-B2, reached single-digit nanomolar affinity for Cn2. Competition experiments suggest they land on the same site as a known neutralizing antibody. That's encouraging. It's also indirect evidence.
The failures carry as much information as the wins. Eleven of 60 single-toxin designs bound, an 18% hit rate that Anthropic compares with a conventional protein design campaign. Every design built to bind more than one toxin at once failed. Ts1 drew zero hits.
Why scorpion stings resist standard treatment
More than 1.2 million people are stung each year, and at least 3,250 die. Most of the deaths are children in places where the nearest hospital is hours away. The World Health Organization lists scorpion sting as a neglected tropical disease that falls hardest on poor rural areas.
Existing antivenom works. Supply is the problem. A product made against North African species won't neutralize a Mexican toxin, and manufacturing still depends on a herd of horses that has to be kept healthy for months before any dose comes out.
That mismatch is what makes the regional split interesting. Seven hits on Cn2 and none on Ts1 is the uneven coverage you'd expect from a first pass at five unrelated toxins.
What the experiment didn't show
No cell-based neutralization test. No animal work.
Not one.
A protein that sticks to a toxin in a dish has cleared the first gate of maybe eight. Binding is necessary and nowhere near sufficient, and Anthropic says outright that it isn't claiming an antivenom.
So what would count as progress? A cell assay in which the toxin stops firing. None of those exist yet.
Outside researchers read the released data the same way. Susana Vázquez Torres at CNIO called the result a promising first step. Timothy P. Jenkins at DTU described it as a case of AI changing how therapeutic discovery gets run, then noted the distance between a binder and a medicine.
That distance is where most programs die. Potency in cells, dosing in animals, safety, manufacturing at volume, and a price per dose that a clinic in a low-income country can pay are separate problems, and each one has killed candidates that looked fine at the binding stage.
What changed since Anthropic's August protein work
The earlier protein design result from Anthropic came from a single prompt and a single pass. This one ran across turns for two and a half days, with the model deciding what to do next each time it finished a step.
For anyone tracking research agents, that's the headline. Generation was never the hard part. Running a campaign across two and a half days is.
Whether the approach travels remains untested. One model, one project, one partner with money.
The designs are public, so academic labs can run the neutralization work Anthropic didn't. Watch for that. If an outside group reports the Cn2 binders blocking the toxin in cells, the 64-hour run stops being a demo and starts being a lead. Until then, this is 17 molecules that bind, plus a long list of things nobody has tried yet.
