Claude's scorpion antivenom test fell far short of a real drug

Claude, the AI model from Anthropic, did not invent a scorpion antivenom. An investment research firm called Citrini Research teamed up with Anthropic to try. Claude produced some promising starting points, but experts told The Verge the result is nothing close to an actual drug.

What the experiment did

Citrini admitted that none of its team knew much about AI as it relates to biology. Anthropic agreed to drop Claude's biosecurity restrictions and give them an unreleased, more advanced model, on the condition that they work from its offices. Anthropic also paid the computing bill, just shy of $50,000. The goal was to design proteins that stick to the toxins in scorpion venom.

Claude worked from a 35-page prompt and some 300 subagents, which are helper copies of the AI. Anthropic said it generated nearly 320,000 candidates. After discarding those that resembled known toxins or existing proteins, 180 designs remained. The team added 31 more with Claude, for 211 in total. The process took 64 hours.

What the lab tests showed

Citrini sent the designs to two outside labs, Adaptyv Bio in Switzerland and Twist Biosciences in the US. That testing cost about $150,000. Seventeen proteins stuck to the toxins. Together they hit four of the five targets. No single protein bound to more than one toxin.

Why experts are not impressed

Sticking to a toxin is not the same as stopping it. Citrini itself said no experiment has yet tested whether the proteins actually stopped the toxin from working. Safety, stability and how the proteins behave in the body are also unknown.

• Andreas Hougaard Laustsen-Kiel of the Technical University of Denmark called finding a good binder "easy-ish." The hard part is turning it into a medicine.
• Two of his Danish colleagues called the findings "far from a medicine and hyped."
• Andreas Bender of Khalifa University said existing methods can already find many hits at manageable cost, even without AI.

The Danish researchers said the novel part was the autonomous discovery process, not the biology.

The business angle

Citrini is not a biotech company. It is the firm behind a widely shared piece of fiction about an AI-driven economic catastrophe that rattled markets. It called the experiment a glimpse of investment opportunities in AI-enabled drug discovery. It did not take the research further, saying it is an investment research firm and not a biotech startup. Anthropic, weeks from its expected IPO, is pursuing drug development and life sciences.

What to keep in mind

AI may well speed up early drug discovery, and the experts say so. But a drug has to work in living bodies and be safe. Until a treatment clears those steps, big claims about AI curing disease are early talk, not results.