Anthropic’s Claude designs protein binders targeting deadly scorpion toxins
AI system generated hundreds of candidate molecules, with lab tests confirming tight binding to a specific Mexican scorpion neurotoxin.
Key points
- Anthropic's Claude autonomously designed protein binders targeting five scorpion toxins.
- Lab testing of 211 candidates confirmed 17 met binding criteria.
- Two strongest binders target Cn2, a toxin from the Mexican scorpion.
Anthropic’s artificial intelligence system, Claude, has independently designed protein binders that target neurotoxins from deadly scorpions, according to a report by The Decoder. The AI operated largely autonomously within Anthropic’s Claude Science research environment to generate molecular structures aimed at neutralizing these specific biological threats.
The system produced 318,484 design candidates for protein binders targeting five different toxins from four distinct scorpion species. After an initial prompt, Claude worked continuously for 24 hours without human intervention. The research team later provided additional instructions to steer the ongoing work and refine the output of the generative process.
Validation and Performance
Researchers selected 180 candidates from the AI’s output and added 31 more designs for a total of 211 molecules tested in the laboratory. Of those tested, 17 met the specific criteria required to be classified as a successful binder. The two strongest performers demonstrated extremely tight binding to Cn2, a specific neurotoxin found in the Mexican scorpion.
Lab tests indicate that these AI-designed molecules dock at or near the same site as an antibody known to neutralize Cn2 in mice. However, the report notes that whether these binders actually function as effective antivenom remains an open question, as such efficacy has not yet been tested in living subjects.
Context and Expert Reaction
Scorpion stings kill more than 3,250 people every year, highlighting the potential significance of new neutralization strategies. The work represents a promising first step, according to external experts Susana Vázquez Torres and Timothy P. Jenkins. Two independent labs confirmed the results, adding credibility to the findings reported by Anthropic’s research team.
This effort builds on an earlier study where Claude designed mini-binders against 14 of 15 target proteins. That previous study reported hit rates of 22 to 35 percent, which compares favorably to the typical 10 to 15 percent hit rates observed in traditional methods. The current scorpion toxin research continues to demonstrate the AI’s capability in complex protein design tasks.
What should you do?
Researchers and biotech professionals should monitor Anthropic’s Claude Science platform for further updates on autonomous protein design capabilities. Institutions involved in toxin research should review the independent lab confirmations to understand the validation standards applied. Stakeholders interested in antivenom development should wait for further efficacy testing before drawing conclusions about clinical viability.
What to do and how to stay safe: Anthropic
- Monitor official announcements from Anthropic regarding the clinical testing phases of these protein binders.
- Review the independent lab confirmations to understand the validation methodology used for the 17 successful binders.
- Assess how autonomous AI design tools like Claude Science might integrate into existing protein engineering workflows.
- Track further studies to determine if the tight binding to Cn2 translates to actual antivenom efficacy in vivo.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Did Claude design the binders completely alone?
Claude worked autonomously for 24 hours after an initial prompt, but the research team later steered the work with additional instructions.
Are these binders confirmed to work as antivenom?
No, whether they actually work as antivenom remains an open question and has not been tested yet.
How many designs were tested in the lab?
Out of more than 318,000 designs generated, 211 were tested in the lab, with 17 meeting the criteria.
