Anthropic: Claude designed binders against scorpion neurotoxins in 64 hours; 17 of 211 lab-tested designs bound, the best at single-digit nanomolar, but multi-toxin designs failed and no neutralization tests yet; all data released

Why it matters

【Key points】

On Oct 20 (13:01 UTC), Anthropic published how Claude designed protein binders against scorpion neurotoxins, with a technical report and full data.

1. With Citrini Research, Anthropic had an internal research model (not named) in Claude Science design binders against five dangerous scorpion toxins (Cn2, Ts1, AaH I, AaH II, LqhIII). Claude chose structures, binding sites and filters; it ran alone for 24 hours, then with four follow-up prompts, 64 hours in all.

2. 12 design tools, 318,484 designs, about 10,000 GPU hours. Claude picked 180, the team added 31; Adaptyv Bio and Twist Bioscience lab-tested all 211.

3. 17 met the bar (hit at both labs, KD of 1 µM or tighter): 7 vs Cn2, 4 AaH II, 4 AaH I, 2 LqhIII, none vs Ts1. Single-toxin hit rate 18%. The best Cn2 binders reach single-digit nanomolar and compete with LR, an antibody known to neutralize Cn2.

4. None of 151 multi-toxin designs worked on more than one toxin, and nothing has been tested for neutralization. Anthropic says this is not an antivenom.

【Verifiable】

Post: https://www.anthropic.com/research/claude-scorpion-toxin-binder-design

Report: https://www-cdn.anthropic.com/31ac578806cffd8b3d92d8affe6edc4c04d502e1.pdf

Data (CC-BY-4.0): https://huggingface.co/datasets/Anthropic/scorpion-toxin-binder-design

【Our read】

1. Another design-to-wet-lab case for Claude, measured by two outside labs with all data public.

2. Designs in days and lab results in weeks matter for neglected diseases like scorpion stings; outside experts at CNIO and DTU call it a promising first step.

3. Multi-toxin binders failed and there is no neutralization data, so a drug is still far off. The partners are research and investment firms, not drug makers.