An AI Program Designs Proteins That Latch Onto Scorpion Venom, and Lab Tests Confirm 17 of Them Work
Anthropic, the company behind the Claude chatbot, reported today that one of its AI programs spent about 64 hours designing small proteins meant to grab onto the poisons in scorpion venom. Two outside labs tested 211 of the designs, and 17 held on firmly. It's an early step toward cheaper antivenom, not an antivenom: nothing has been tested in animals or people yet.
Anthropic, the company that makes the Claude chatbot, published a report today on an experiment in designing medicine-like molecules with AI. The goal was proteins that latch onto the poisons in scorpion venom, the first step toward a new kind of antivenom.
Scorpion stings are a bigger problem than many people realize. More than 1.2 million people are stung each year and more than 3,250 die, and health experts class scorpion stings as a neglected tropical disease, one that gets little attention or money. Today's antivenoms are made by injecting horses with small doses of venom and collecting the antibodies their bodies make, which takes months and is costly. They are often expensive or in short supply where they're needed.
Here is what the company did. It gave an internal version of Claude, one the public can't use, five poisons from the venom of scorpions in Mexico, Brazil, North Africa and the Middle East, plus a dozen free protein-design programs. Claude then worked largely on its own for about 64 hours. Researchers stepped in four times, mostly to push it toward designs that could grab several poisons at once. It produced 318,484 candidate proteins, checked on the computer how each would fit against each poison, and picked 180 to make for real. The team added 31 more, for 211 in all. Two outside companies, Adaptyv Bio and Twist Bioscience, built and tested them.
Seventeen of the 211 held firmly onto their target. They covered four of the five poisons; none worked against the one from Brazil. The two best grabbed the Mexican scorpion's poison very tightly, at or near the same spot as an antibody that is known to protect mice from it. The idea came from Citrini Research, a financial research firm, and Anthropic has released all 211 designs and the test results free for anyone to use.
The company is clear about what this isn't. The proteins were only tested for whether they stick to the poison. No one has yet checked whether they stop it from doing harm, in living cells or in animals, and Anthropic says it does "not claim to have produced an antivenom." The harder goal, one protein that grabs several scorpions' poisons at once, failed completely. Two outside scientists who reviewed the work called it a promising proof of concept. One of them, Susana Vázquez Torres of Spain's National Cancer Research Centre, said the real challenge is turning tools like this into "molecules that can become real drugs."