An experiment involving Anthropic’s Claude produced dozens of protein designs that bound to scorpion toxins in laboratory tests. That is a meaningful early research result, but it does not mean the team created an antivenom: no experiment tested whether the proteins stopped toxins from working, and much remains unknown about their safety and behavior in the body.
From candidate designs to laboratory tests
Citrini Research, an investment research firm, set out to explore whether AI could help develop a new drug. It chose scorpion antivenom, a treatment that can be costly and difficult to produce, often involving injecting venom into horses. The research goal was to design proteins that could bind to toxins in scorpion venom.
After Claude’s biosecurity safeguards blocked the firm’s initial attempts to research drug candidates, Citrini connected with Anthropic. The company provided access to an unreleased model and covered computing and token costs that came to just shy of $50,000. The team worked from Anthropic’s offices.
With a 35-page prompt, the team used roughly 300 subagents to search for suitable proteins. Anthropic’s account says Claude generated nearly 320,000 candidates and narrowed them to 180 designs after removing candidates resembling known toxins or existing proteins. The process took 64 hours. The team then worked “with Claude” to add 31 more designs, for a total of 211.
Citrini sent the designs to Swiss biotech company Adaptyv Bio and US-based Twist Biosciences for testing. Together, 17 proteins bound to four of the five scorpion toxins the team aimed to target. Each protein bound to only one toxin. The testing cost another $150,000, give or take.
Binding is an early milestone, not a treatment
A protein binding to a toxin is only one step in developing a medicine. It does not establish that the protein can neutralize the toxin, work safely, or function as a treatment. Citrini acknowledged that “no experiment has yet tested whether they actually stopped the toxin working.”
That gap is why the experiment’s results should be read as early candidate discovery rather than drug development completed. The proteins still need to be understood and tested beyond whether they attach to their intended targets. Their safety, stability, behavior in the body, and potential to become viable treatments remain uncertain.
Researchers at the Technical University of Denmark said the findings were technically promising, but cautioned that the excitement was premature. Andreas Hougaard Laustsen-Kiel, a professor of antibody technologies there, described finding a binder as the easier part; translating it into medicine is the harder work. Melisa Benard Valle, an assistant professor, and Anne Ljungars, a senior researcher, likewise said that the path from these designs to a medicine is likely to be long and challenging.
The larger question is what AI changes
For the researchers, the experiment’s most novel feature may have been how the discovery process was carried out, rather than the biological results themselves. AI may help accelerate the search for candidate molecules. But finding a promising molecule and showing that it works in living systems are different problems.
Andreas Bender, a professor of machine learning in medicine at Khalifa University in Abu Dhabi and visiting professor at the University of Cambridge, said existing techniques such as high-throughput screening can find many candidate hits at manageable cost without AI. Designing proteins that bind to a target is, in his view, “kind of” a solved problem. The missing challenge is translating those results into effective treatments.
That distinction matters when assessing claims about AI drug discovery. A large number of generated candidates, followed by a smaller group that binds in a test, can show that a system contributed to an early-stage workflow. It does not, by itself, establish that AI has shortened the path to a usable drug.
A research demonstration with an investment angle
Citrini called the project a success of sorts and presented it as a glimpse of investment opportunities in AI-enabled drug discovery. The firm said it was an investment research firm, not a biotech startup, when responding to criticism that it had not taken the work further.
The experiment also involved substantial costs: Anthropic covered nearly $50,000 for computing and tokens, while outside laboratory testing cost another $150,000, give or take. Bloomberg reported that Citrini highlighted companies involved in drug development testing, including Charles River Laboratories. The article noted that Citrini described the AI portion as “free” in a video, despite Anthropic covering its compute bill.
The evidence supports a narrower conclusion than “Claude made an antivenom.” The model helped generate protein candidates, and some bound to targeted toxins in tests. Whether any could neutralize those toxins or lead to a safe, effective treatment remains unanswered.