Anthropic has announced that an internal version of its Claude AI model operating within Claude Science largely autonomously designed novel protein minibinders against scorpion venom neurotoxins. Tested across two independent contract research laboratories, 17 computationally generated designs demonstrated verified laboratory binding against four medically dangerous neurotoxin targets. The experiment originated from a collaboration with investment research firm Citrini Research to examine whether frontier artificial intelligence could enable small teams to tackle neglected global health challenges.
Scorpion envenomation Causes an estimated 1.2 million stings and over 3,250 deaths annually, disproportionately affecting children in underserved tropical regions where access to healthcare is limited. Conventional antivenom production requires immunizing horses with milked scorpion venom, a costly and slow process that yields supplies with limited regional utility. To test an alternative de novo approach, the research team tasked Claude with designing small computational protein binders capable of latching onto the channel-binding interfaces of five scorpion neurotoxins spanning Mexico, Brazil, North Africa, and the Middle East: Cn2, Ts1, AaH I, AaH II, and LqhIII.
Autonomous Computational Campaign Across 64 Hours
Working under human supervision with four follow-up prompts, the internal AI research model operated across 12 protein design and folding tools over 64 hours, consuming approximately 10,000 GPU hours. Claude selected candidate target structures from the Protein Data Bank and AlphaFold database, identified binding hotspots, and determined scoring thresholds. In total, the system generated 318,484 de novo designs, co-folded and screened 553,052 design-toxin complexes, and picked a priority pool of 180 designs. An additional 31 designs were added with the assistance of researchers, resulting in 211 candidates forwarded for physical laboratory synthesis.
Independent Laboratories Confirm 17 Binding Hits
Contract research organizations Adaptyv Bio and Twist Bioscience expressed the engineered proteins using cell-free bacterial systems and mammalian cell cultures, respectively. Binding was measured via surface plasmon resonance to assess dissociation constants and establish affinity. Seventeen designs met the evaluation criteria, each demonstrating high-affinity binding to a single toxin target while showing no off-target binding to an unrelated human control protein. Among the candidates directed at single toxins, 11 of 60 succeeded, establishing an 18 percent hit rate consistent with conventional protein design benchmarks.
Early Molecular Progress Faces Crucial Downstream Hurdles
The two most potent binders, designated Cn2-B1 and Cn2-B2, targeted the Mexican scorpion toxin Cn2 with affinities in the single-digit nanomolar range. In competition experiments, these candidates bound the same site as an existing antibody known to neutralize Cn2 in animal models, suggesting potential therapeutic relevance. However, none of the candidates designed to target multiple toxins simultaneously proved successful across more than one target. Anthropic noted that the binders were evaluated only against lab-produced recombinant toxins rather than whole natural venom, with cellular neutralization and live animal protection studies still required to determine whether the molecules can form the basis of a clinical antivenom.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
