Claude designed a protein to combat scorpion venom

Summary

Anthropic's Claude model autonomously designed miniproteins capable of binding to scorpion neurotoxins. Out of 211 candidates, 17 met the binding criteria, with two showing nanomolar affinity against the Centruroides noxius toxin. Conducted at the request of Citrini Research, this experiment tested whether artificial intelligence can accelerate early-stage biomedicinal design for a problem affecting over 1.2 million people annually.

Why it matters

Scorpion stings cause at least 3,250 deaths each year. Most of these fatalities occur among children in regions with limited access to antivenoms, which are expensive due to waste-derived sources and months-long production processes. If AI can reduce the protein design phase from months to days, it could help spread early-stage antivenom research in low-resource settings. Additionally, the open-source publication of all 211 designs lowers the barrier for independent labs to build upon this work.

Highlights

• Claude operated autonomously for 64 hours, using 12 expert tools to generate 318,484 candidates and narrow them down to 211 testable designs.
• Two top-tier proteins showed nanomolar-level affinity and bound to the neurotoxin of a Mexican scorpion. However, it has not yet been proven that this binding is equivalent to functional neutralization.
• All 211 protein designs and associated data were openly published to allow the research community to build upon this work.

Critical take

The gap between experimental binding and actual venom neutralization in a living organism is vast. The team acknowledges that protection at the cellular or animal level has not yet been demonstrated. Presenting this finding as a step toward antivenoms risks overstating a study that is still essentially a binding test. Furthermore, since proteins binding to multiple toxins have not been produced, the claim of a 'broad-spectrum antivenom' remains a goal rather than an achieved result.