The short version
- An unreleased version of Claude helped generate over two hundred protein designs intended to neutralize scorpion venom toxins.
- Laboratory testing confirmed that seventeen candidates bound to specific toxins, but none demonstrated the ability to neutralize them or function as medicine.
- Biologists caution that while the autonomous discovery process is novel, translating these binders into safe and effective drugs remains a significant hurdle.
A recent experiment involving an unreleased version of Anthropic’s Claude AI model has clarified the current limits of artificial intelligence in drug development. The project, initiated by investment research firm Citrini Research, aimed to demonstrate whether AI could revolutionize the creation of scorpion antivenom. While the collaboration produced a set of protein designs that successfully attached to specific toxins, independent experts emphasize that these results do not constitute a new drug or even a viable treatment candidate.
The initiative began after Citrini researchers encountered biosecurity restrictions while attempting to use AI for drug research. Seeking to bypass these guardrails, the firm contacted Anthropic, which agreed to provide access to an advanced model and cover computing costs in exchange for on-site collaboration. The partnership resulted in a blog post detailing the process, though the scientific community has responded with caution regarding the claims made about the experiment’s success.
The technical workflow involved submitting a detailed prompt to Claude, which then deployed hundreds of subagents to identify potential protein structures. Over a period of sixty-four hours, the system generated nearly three hundred thousand candidates. After filtering out designs that resembled known toxins or existing proteins, the team narrowed the selection to one hundred eighty options. Researchers subsequently added thirty-one more designs through direct interaction with the model, bringing the total to two hundred eleven candidates for further evaluation.
To assess the viability of these designs, Citrini sent them to external laboratories, including Swiss biotech firm Adaptyv Bio and US-based Twist Biosciences. This testing phase incurred approximately one hundred fifty thousand dollars in costs. The results showed that seventeen of the proteins successfully bound to four of the five targeted scorpion toxins. However, none of the candidates bound to more than a single toxin, and crucially, no experiment was conducted to determine whether these bindings actually neutralized the venom’s effects.
Experts in antibody technology have pointed out that binding to a toxin is only an initial step in drug development. Andreas Hougaard Laustsen-Kiel, a professor at the Technical University of Denmark, noted that discovering binders for snake and scorpion toxins is relatively straightforward compared to the complex process of translating those binders into safe and effective medicines. The current results represent what some describe as the easier phase of discovery, leaving significant challenges ahead.
Melisa Benard Valle and Anne Ljungars, also from the Technical University of Denmark, described the findings as technically promising but cautioned against premature excitement. They highlighted that much remains unknown about the safety, stability, and biological behavior of the AI-designed proteins. According to their assessment, the path from these initial binders to a functional medicine is likely to be long and difficult, requiring extensive further research and development.
The collaboration has drawn attention partly due to Citrini’s previous involvement in circulating speculative narratives about economic impacts of AI. The firm acknowledged its lack of expertise in AI-related biology at the outset of the project. Anthropic, approaching a highly anticipated public offering, has been actively promoting the potential of AI in life sciences, making this partnership a notable example of industry efforts to showcase practical applications.
While the experiment did not produce a new antivenom, it did demonstrate a novel approach to autonomous molecular discovery. The ability of AI to rapidly generate and filter large numbers of candidate structures offers insights into how technology might accelerate early-stage research. However, the gap between computational design and clinical reality remains wide, underscoring the need for rigorous biological validation in future studies.
The outcome serves as a reminder that while artificial intelligence can enhance certain aspects of scientific inquiry, it does not replace the complex, multi-step process of drug development. The proteins identified in this study require further investigation to determine their potential utility. Until then, they remain theoretical candidates rather than practical solutions for treating scorpion envenomation.
As AI continues to penetrate various sectors of healthcare and biotechnology, collaborations like this one will likely become more common. Stakeholders must balance optimism about technological capabilities with realistic expectations regarding timelines and outcomes. The distinction between generating data and creating medicine is critical, and this experiment illustrates that distinction clearly.
Sources behind this briefing
Go to the original reporting
- The Verge↗Claude did not invent scorpion antivenom
