Citrini 团队用 Anthropic Claude 自主设计蝎子毒素结合蛋白，实验室验证部分结合成功

In a test of artificial intelligence's ( AIQ ) ( AIEQ ) capabilities in drug discovery, a team of financial analysts with virtually no scientific background used Anthropic's ( ANTHRO ) Claude to autonomously design protein molecules that successfully bound to scorpion toxins in laboratory testing. The experiment Citrini Research budgeted $200K to determine whether non-scientists can use AI to create something that actually works in a test tube. Working on-site at Anthropic's ( ANTHRO ) offices with an advanced internal research model, the five-person team—three with no science background and two with minimal experience—prompted Claude to design minibinders targeting scorpion toxins. The AI spawned more than 300 sub-agents using over 300 concurrent GPUs, completing the initial design campaign in 64 hours. The level of autonomy proved striking. Claude searched scientific literature, coordinated specialist software, managed computational resources, and even drafted and sent emails to scientists at Twist Bioscience to resolve manufacturing questions—all while the human team was away. The entire design process that produced 180 initial candidates required just five messages from humans to Claude. Laboratory results Contract research organizations Adaptyv Bio and Twist Bioscience synthesized 211 AI-designed proteins and tested them using Surface Plasmon Resonance. The results exceeded expectations: 17 designs successfully bound to their targets across four of five toxins—without any laboratory feedback to improve them. Two designs, Cn2-B1 and Cn2-B2, achieved tight binding in the low-nanomolar range, comparable to known antibodies. The "moonshot" attempt to create a single molecule binding all five toxins failed, as Claude had predicted. No experiment has yet confirmed whether the binders actually neutralize toxin activity, and full drug development would require extensive animal testing and human trials. Still, as Citrini concluded : "The frontier has moved." Their experience suggests AI in biology will progress "in a manner that is both alarmingly fast and not fast enough"—compressing years of design work into days while remaining bound by the irreducible pace of physical science.