Anthropic's Claude AI system designed hundreds of thousands of candidate proteins in 64 hours to target scorpion venom toxins, with 17 designs meeting the company's binding criteria in laboratory tests. The results offer an early glimpse of how AI could accelerate drug discovery, though the proteins have not been shown to neutralize venom.
In a research report, Anthropic outlined an experiment conducted with input from investment research firm Citrini Research and validated through laboratory testing by Adaptyv Bio and Twist Bioscience. The company is releasing the sequences and experimental measurements for all 211 designs tested, allowing other researchers to build on the findings.
Anthropic focused on scorpion envenomation, which it said causes at least 1.2 million stings and more than 3,250 deaths annually. Children in regions with limited access to medical care face particular risks. The company also cited the high cost of developing new treatments and limited financial incentives as barriers to addressing the problem.
How Claude Designed The Proteins
The experiment tasked Claude with designing proteins, known as minibinders, that could attach to five toxins associated with scorpions in Mexico, Brazil, North Africa and the Middle East.
The system used structural information from the Protein Data Bank and AlphaFold's database to identify toxin regions involved in interactions with sodium channels, which help nerve cells transmit electrical signals.
Anthropic said Claude generated hundreds of thousands of candidates in 64 hours before the team narrowed the designs to 180 for laboratory testing. Researchers subsequently added 31 more candidates, bringing the total to 211.
Of those, 17 met the team's threshold for binding to their intended toxins. However, none successfully targeted multiple toxins, with all 17 qualifying designs binding to just one.
An Early Step, Not An Antivenom
Anthropic cautioned that the experiment has not established whether the proteins can neutralize venom in cells, much less treat a scorpion sting. Binding to a toxin is an important early milestone, but it does not establish that a candidate can prevent the toxin from causing harm.
External reviewers Susana Vázquez Torres and Timothy P. Jenkins viewed the results as an encouraging starting point.
"Designing credible binders against five toxins in parallel, in a matter of days, is a great showcase of the AI-driven paradigm shift we are seeing in therapeutic discovery," Jenkins said.
Traditional antivenom production typically involves collecting venom, immunizing animals and purifying the resulting antibodies, a process Anthropic described as resource-intensive and slow.
Computationally designed proteins could eventually be refined for greater stability and potency, potentially opening another route to developing treatments. But Anthropic acknowledged that the more difficult work of turning promising binders into effective therapies remains ahead.
