- Claude generated over 318,000 protein designs in 64 hours, with 17 passing tests.
- Scorpion stings affect at least 1.2 million people annually, resulting in over 3,250 deaths.
- The research focuses on five scorpion neurotoxins from different regions.
- Further testing will examine toxin neutralization, safety, and therapeutic efficacy.
In just 64 hours, Claude generated over 318,000 protein designs, with 17 passing laboratory binding tests against dangerous scorpion neurotoxins, a move that could accelerate antivenom research. The study was supported by Citrini Research and conducted by Adaptyv Bio and Twist Bioscience in laboratory tests. A total of 211 selected designs were tested, with 17 meeting strict binding criteria.
Early Steps in Antivenom Development
The strongest candidates achieved binding affinities at the nanomolar level for Mexican scorpion venom, indicating tight molecular interactions. These results mark early steps in the potential development of antivenoms, rather than ready-to-use therapeutic solutions.
Each year, scorpion stings affect at least 1.2 million people, leading to over 3,250 deaths, primarily in areas with limited medical resources. Existing antivenoms are typically produced by immunizing animals, including horses, to extract antibodies, a process that can take months and is effective only against a limited range of toxins. To address these limitations, Anthropic tasked its internal Claude model to design small binding proteins (minibinders), which are specifically engineered to attach to targeted molecules.
The project focuses on five scorpion neurotoxins from Mexico, Brazil, North Africa, and the Middle East.
Binding Affinity Testing of Candidates
Claude primarily operated autonomously, generating 318,484 candidates using 12 protein design tools and calculating 553,052 pairs of protein-toxin combinations. The project consumed approximately 10,000 GPU hours. After initial prompts and four follow-up prompts, Claude selected 180 designs for laboratory testing. Researchers subsequently added 31 more designs with the model's assistance. The final 211 candidates included proteins specifically targeting single toxins as well as proteins designed to bind multiple toxins simultaneously.
Adaptyv Bio and Twist Bioscience employed different laboratory methods to produce these proteins and measured binding levels using surface plasmon resonance technology, which tracks the strength of molecular attachment and the speed of their separation.
To qualify, each candidate had to demonstrate stable binding in two laboratories, meet specified affinity thresholds, and show no detectable binding to an irrelevant human protein used as a control. Ultimately, 17 designs met these criteria, including 7 targeting Mexican scorpion toxin Cn2, 4 targeting AaH I and AaH II, and 2 targeting LqhIII. All multi-toxin designs failed to target more than one toxin. Among the 60 designs developed specifically for single toxins, 11 were successful, yielding an approximately 18% hit rate.
The two main candidates, Cn2-B1 and Cn2-B2, exhibited particularly strong binding affinities. Competitive experiments indicated that their attachment sites were similar to those of antibodies previously shown to neutralize Cn2 in mice. However, researchers only conducted binding tests on laboratory-produced toxins and did not test those extracted from scorpions, leaving it unclear whether these candidates could inhibit toxin activity in cells or protect animals.
Further testing is needed to examine toxin neutralization, safety, stability, and therapeutic efficacy before considering clinical development. Anthropic is releasing all 211 designs, predicted structures, and experimental results to support further research. The company detailed its findings in a technical report rather than publishing in a peer-reviewed journal.
Claude's Contribution to Antivenom Research
Claude's research offers new insights into the development of antivenoms, especially as existing treatments face challenges. A significant number of people are affected by scorpion stings each year, making rapid and effective antivenom development crucial. The protein designs generated by Claude are not only numerous but also precisely tailored to specific toxins, potentially paving the way for future therapeutic solutions. While further testing is still required to determine their actual efficacy, this research demonstrates the potential of artificial intelligence in the biomedical field.