Claude Autonomously Designs 17 Scorpion-Toxin Binders in 64 Hours
TL;DR
- Citrini Research, a 5-person investment firm with no biology background, ran the full experiment for under $200K including $49,997 in compute costs.
- Claude used a 35-page prompt and 300+ coordinated agents to generate 318,484 candidates over more than 10,000 H100 GPU hours.
- All 211 designs and binding data were published on Hugging Face rather than submitted to peer review, bypassing traditional pre-publication validation.
Anthropic says Claude ran a 64-hour autonomous protein-design campaign off one initial prompt and four follow-ups, and produced 17 binders to scorpion neurotoxins, two of them at single-digit nanomolar affinity. Across the run the model generated 318,484 designs, screened 553,052 design-toxin pairs, and used 12 tools, according to Anthropic's research post. Claude picked 180 candidates for lab work; the team added 31 more, for 211 total tested at Adaptyv Bio and Twist Bioscience. The compute bill came to roughly 10,000 GPU hours.
The targets were five neurotoxins from scorpions in Mexico, Brazil, North Africa and the Middle East: Cn2, Ts1, AaH I, AaH II and LqhIII. Claude hit 11 of 60 single-toxin binders tested, an 18% success rate. Multi-toxin binders met the criteria zero times. The two strongest, Cn2-B1 and Cn2-B2, had dissociation constants "in the single-digit nanomolar range", well below the 1 µM threshold the team set.
Then the caveat Anthropic itself leads with: "We have not confirmed whether these binders neutralize scorpion venom in a cellular environment, so we do not claim to have produced an antivenom." The post calls out two unresolved steps, cell-based neutralization and animal efficacy, and notes the binders were tested only against recombinant toxins rather than venom-purified ones.
Two outside collaborators carry the voice on the result. Susana Vázquez Torres, a postdoctoral researcher at the Spanish National Cancer Research Centre, said: "Designing a binder against almost any target no longer seems unrealistic. We can now obtain very tight binders remarkably quickly, and binding itself is no longer the main bottleneck." Timothy P. Jenkins, an associate professor at the Technical University of Denmark, called the run "a great showcase of the AI-driven paradigm shift we are seeing in therapeutic discovery."
Scorpion envenomation kills more than 3,250 people a year out of 1.2 million-plus annual stings, a neglected tropical disease. Citrini Research is listed as a collaborator alongside the two CROs.
The post lands in a visible run of agent-driven drug-discovery coverage, including OpenAI's purchase of Orchestra Bio earlier the same day. But this one lives or dies on the next data drop: whether any of the 17 binders actually neutralizes the toxin.
What others are reporting
-
Hunterbrook Media Read →
Investigative; CEO participated directly; FDA expert flags unresolved clinical development questions; explicit disclosure of long/short AI biotech positions.
AI is supercharging drug discovery. But there's still questions around capital, safety, and the right experiments to do.
-
Interesting Engineering Read →
Global health framing (1.2M stings, 3,250 deaths/yr); details both labs' different production methods; notes open Hugging Face release of all data.
-
GIGAZINE Read →
Emphasizes Anthropic's explicit disclaimer that binding does not equal antivenom; cites 10,000 GPU hours consumed and the public Hugging Face data release.
if AI can design binding molecules for hard-to-target toxins in days and verify them in the laboratory within weeks, it could theoretically accelerate the discovery of treatments
-
36Kr Read →
Chinese business press; breaks down total cost to ~$200K ($49,997 compute); reveals 35-page prompt, 300+ agents, and 5-person non-biology team composition.
211 candidate molecules were sent to the laboratory, 17 passed the binding test, covering 4 types of high-risk scorpion toxins.
-
DEV Community Read →
Critical developer framing; argues design speed does not equal development timeline; flags that Claude's full toolset is not broadly accessible today.
Binding is only one property a potential therapeutic or diagnostic molecule may need.
Originally reported by anthropic.com
Read the original article →Original headline: Anthropic Says Claude Autonomously Designed 17 Scorpion-Toxin Antivenom Binders in a 64-Hour Campaign