The team at Citrini Research worked with Anthropic to see if motivated individuals with no lab experience could use AI to design a working drug For years, artificial-intelligence evangelists have boasted about the technology’s capacity to revolutionize drug development and medical research. One team of independent investment researchers recently decided to put those claims to the test. Citrini Research, an independent research shop and investing newsletter founded by James Van Geelen, set out to see whether a team of motivated individuals with little to no lab experience could use Claude’s frontier AI models to design novel drugs that work when put to the test.
After being stonewalled by automated model guardrails, the Citrini team reached out to Anthropic. They eventually received permission to carry out their experiment on-site at Anthropic’s offices using an internal, unreleased Claude Science research model. After brainstorming suitable targets with Claude, the team settled on designing a synthetic protein that could serve as an antidote to scorpion venom.
Why humanoid robots are so hard to mass-produce Play video: Why humanoid robots are so hard to mass-produce Existing scorpion antivenoms are produced by injecting horses with venom and harvesting their plasma antibodies, making them expensive to produce, variable and effective only against specific species. More than 3,200 people die every year from scorpion stings globally, according to figures cited in the report. Don’t Short Yourself offers weekly money tips to help you earn it, stack it and grow it.
I would like to receive updates and special offers from Dow Jones and affiliates. I can unsubscribe at any time. The team set a moonshot goal: developing a single, universal protein “minibinder” that could neutralize five dominant scorpion neurotoxins across different species.
While the team ultimately failed to achieve that universal moonshot goal, the broader experiment still managed to produce 17 working proteins that successfully bound to four of the targeted individual toxins. Those 17 proteins were among 211 candidate designs generated by Claude that were sent to real-world research laboratories for physical testing. The total budget for the experiment — including AI computing, tokens and wet-lab synthesis — was approximately $200,000, Citrini said.
The team used Claude to help prepare a 35-page prompt that launched the design run. The initial computational design phase took just 64 hours and required only five messages to Claude. The primary takeaway, according to Citrini, is that while AI-enabled research can dramatically compress the intellectual and design phase of drug development down to a few days, severe physical bottlenecks remain.
Extract — continue reading at the source.
