Citrini Research publishes its account of the project, noting five total human messages and a 64-hour initial campaign producing the best design Cn2-B1
Claude Autonomously Designed 17 Confirmed Scorpion Neurotoxin Binders
The result demonstrates an AI system autonomously completing a multi-step computational protein design campaign with minimal human input and producing lab-confirmed binders. Scorpions sting approximately 1.2 million people each year, and current antivenoms are often costly or in short supply, according to Anthropic.
The full picture
Anthropic published findings describing Claude running an autonomous protein design campaign against scorpion neurotoxins. Claude selected target sites, drove 12 design programs, set its own scoring criteria, and generated 318,484 candidate molecules. It selected 180 for physical synthesis; Anthropic's team added 31 more, bringing the total to 211 designs sent to lab partners Adaptyv Bio and Twist Bioscience for testing. Of those, 17 were confirmed to bind their target toxins. The project was commissioned by Citrini Research, described as a finance firm with no lab or protein expertise, and the entire original design process involved only five messages sent by humans to Claude, completing within 64 hours. The best-performing design, Cn2-B1, was produced during the third follow-up prompt. Testing was conducted against recombinant toxins, not purified venom, and no neutralization or in-vivo efficacy tests were performed. None of the designs intended to bind two or more toxins met the study's criteria against more than one target, and no qualifying binders were found for Ts1, a medically important Brazilian scorpion toxin. Anthropic is releasing all 211 designs and their binding data publicly.
How it developed
Sources
1 more source
Want this in your inbox?
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free