Claude 并未造出蝎毒抗毒血清，Citrini 与 Anthropic 的制药实验被指夸大

An investment research firm wanted to prove AI could revolutionize drug development. So they teamed up with Anthropic to see whether an unreleased version of Claude could make scorpion antivenom. It couldn’t. It came up with some promising starting points, but nothing close to an actual drug, experts tell The Verge.

Citrini Research, the firm behind the experiment, admitted that “none of us could claim to know anything about AI as it related to biology.” That didn’t stop it from trying to make a drug from scratch — or publishing a blog post titled “We Wanted to Use Claude to Create a New Drug. Here’s What Happened.” The account came complete with rhetorical questions like “Why don’t we try to use AI to create a new drug?”, a blow-by-blow account of its amateurish scientific adventures, and a quote from Charles Darwin.

If the name Citrini sounds familiar, that’s because the firm was behind a widely circulated piece of fiction imagining an AI-induced economic catastrophe that gained traction on social media and rattled real-world markets. It also claims to have visited the Strait of Hormuz in April to get a firsthand perspective on what was happening in the volatile region.

For its latest venture, Citrini enlisted the help of Anthropic, which, mere weeks out from its hotly anticipated IPO, has a few reasons to talk up AI’s potential in drug development and the life sciences — two areas it’s actively pursuing.

The collaboration came about after Citrini ran into Claude’s biosecurity guardrails while trying to get the model to research drug candidates. The post’s author said they “spam called anyone in my contact list who worked at an AI lab,” eventually finding a willing partner in Anthropic. The AI lab agreed to drop the restrictions and give the team access to a more advanced, unreleased model on the condition that they work from its offices. “Luckily,” Anthropic also agreed to foot the hefty compute and token bill, which came just shy of $50,000.

After weighing a few options, the team settled on scorpion antivenom, a costly treatment that can be difficult to produce, often involving injecting venom into horses. The goal was to use Claude to design proteins that could bind to the toxins in scorpion venom. Armed with a 35-page prompt, they set the AI to work, unleashing some 300 subagents to discover suitable proteins.

In its own more sober account of the experiment, Anthropic said Claude generated nearly 320,000 candidates, whittling them down to 180 designs after discarding those that resembled known toxins or existing proteins. The entire process took 64 hours. Anthropic said the team worked “with Claude” to add another 31 designs, bringing the total to 211.

Citrini then sent the designs to two outside laboratories — Swiss biotech company Adaptyv Bio and US-based Twist Biosciences — to test whether the proteins could actually bind to their intended targets, a key step for developing an effective antivenom. That cost another $150,000, give or take. It’s unclear why Anthropic could not have conducted these tests in-house at its biology wet lab in California. Seventeen of the proteins passed muster, collectively binding to four of the five scorpion toxins the team had hoped to target. Not one of the proteins bound to more than a single toxin, Citrini said.

Even if they had, Citrini would still be a very long way from a working antivenom. Binding to a toxin does not mean a protein can actually neutralize it, let alone function as a safe and effective treatment. Citrini acknowledges this itself, noting that “no experiment has yet tested whether they actually stopped the toxin working.”

In other words, the experiment had not succeeded. Citrini set out to make a new drug, but experts tell The Verge it had only accomplished the relatively “easy” part of the process. “It is easy-ish to discover a good binder against snake and scorpion toxins,” said Andreas Hougaard Laustsen-Kiel, a professor of antibody technologies at the Technical University of Denmark. “The hard part is to translate it into a medicine.”

Melisa Benard Valle, an assistant professor, and Anne Ljungars, a senior researcher, both at the Technical University of Denmark, agreed. “The results are technically promising, but some of the excitement is premature,” they wrote in a joint email, describing the findings as “far from a medicine and hyped.” Much remains unknown about the proteins Claude designed, they said, including their safety, stability, how they behave in the body, and whether they can be developed into viable treatments. “Translating these binders into an actual medicine is likely still a long and challenging process.”

The most novel aspect of Citrini’s experiment, they said, was the “autonomous discovery process rather than the biological results.” AI has huge potential to accelerate discovery, they added, but warned “turning promising molecules into effective medicines remains a much greater challenge.”

Nor is the underlying approach particularly new. Andreas Bender, a professor of machine learning in medicine at Khalifa University in Abu Dhabi, told The Verge existing methods like high-throughput screening “are often able to find a multitude of hits at manageable cost,” even without any AI. Designing proteins that bind to targets is “kind of” a solved problem, said Bender, who is also a visiting professor at the University of Cambridge. The much harder challenge is turning those results into treatments that actually work in living systems. “It’s exactly this translation that lacks.”

While it failed to make a drug, Citrini declared the experiment a success of sorts. The firm argued it offers a glimpse into the investment opportunities emerging in the upcoming era of AI-enabled drug discovery. As for why it didn’t bother to take the research further? Citrini accused critics of “nitpick[ing]” and added, “Not to beat a dead horse, but we are an investment research firm - not a biotech startup.” Citrini and Anthropic did not respond to requests for comment.

But it still made a bullish case for companies that might benefit. Naturally, the specifics were reserved for paying subscribers, though Bloomberg reports the firm highlighted companies involved in the testing side of drug development, including Charles River Laboratories, which supplies primates for animal research. In a video discussing the experiment, Citrini described the AI portion as “free,” conveniently overlooking the tab picked up by Anthropic.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

• Robert Hart