No science background? AI does the lab work
In a test of artificial intelligence's AIQ AIEQ ability to discover new drugs, a team of financial analysts with almost no science experience used Anthropic's ANTHROPIC Claude AI to design protein molecules that successfully attached to scorpion toxins in laboratory tests.
The experiment
Citrini Research set aside $200K to find out whether people without science training could use AI to create something that actually works in a test tube.
The five-person team worked at Anthropic's ANTHROPIC offices and used an advanced research model that was not publicly available. Three team members had no science background, and two had only a little experience. They asked Claude to design small proteins, called minibinders, that could attach to scorpion toxins. The AI created more than 300 smaller AI agents and used more than 300 GPUs at the same time. It finished the first round of designs in 64 hours.
The AI worked with a surprising amount of independence. Claude searched scientific papers, coordinated specialized software, managed computing resources, and even wrote and sent emails to scientists at Twist Bioscience to get answers about how to make the proteins. It did all this while the human team was away. The entire design process that produced 180 first-round candidates took only five messages from the humans to Claude.
Laboratory results
Two research companies, Adaptyv Bio and Twist Bioscience, made 211 proteins designed by AI and tested them using a method called Surface Plasmon Resonance.
The results were better than expected. Seventeen designs successfully attached to their targets among four of the five toxins. The AI had received no laboratory test results to help it improve its designs. Two designs, Cn2-B1 and Cn2-B2, attached very strongly, with binding strengths in the low-nanomolar range. This was similar to the strength of known antibodies.
The more ambitious attempt to create one molecule that could attach to all five toxins did not work, just as Claude had predicted.
So far, no experiment has shown whether these proteins can actually stop the toxins from working. Developing a complete drug would also require extensive testing in animals and clinical trials in humans.
Even so, Citrini concluded: "The frontier has moved." The team's experience suggests that AI in biology will advance "in a manner that is both alarmingly fast and not fast enough". AI can reduce years of design work to days, but scientific progress still depends on physical experiments, which take time that cannot simply be removed.