Claude designed protein binders against scorpion neurotoxins
anthropic.com
October 20, 2026
12 min read
49/100
Summary
Anthropic used an internal Claude research model to design small protein binders for five scorpion-venom neurotoxins, with the long-term goal of supporting antivenom development. Over 64 hours and four follow-up prompts, Claude generated 318,484 candidate proteins, screened 553,052 protein-toxin pairs, and selected 180 designs for laboratory testing. Anthropic added 31 more designs with Claude’s help, and Adaptyv Bio and Twist Bioscience tested all 211 candidates. Seventeen designs met Anthropic’s binding criteria, which required consistent results from both laboratories, dissociation constants of 1 µM or tighter, and no detectable binding to an unrelated human protein. The binders each targeted a single toxin: seven bound Cn2, four bound AaH II, four bound AaH I, and two bound LqhIII. No designs intended to bind multiple toxins met the criteria for more than one target. The strongest Cn2 binders, Cn2-B1 and Cn2-B2, showed single-digit nanomolar affinity and competed with an antibody reported to neutralize Cn2 in mice. The results establish binding to recombinant toxins, not venom neutralization. Anthropic has not tested the designs in cells or animals and does not claim to have produced an antivenom. It released all 211 designs and their experimental data for further research.
Key Takeaways
- Claude generated 318,484 protein-binder designs against five scorpion neurotoxins and selected 180 for testing after a 64-hour campaign using roughly 10,000 GPU hours.
- Adaptyv Bio and Twist Bioscience tested 211 designs, and 17 met Anthropic’s criteria for toxin binding across both laboratories.
- The 17 successful binders targeted individual toxins; no multitoxin design met the binding criteria for more than one target.
- Cn2-B1 and Cn2-B2 bound the Mexican scorpion toxin Cn2 with single-digit nanomolar affinity and competed with a Cn2-neutralizing antibody for a similar binding site.
- Anthropic has not tested whether any binder neutralizes toxin activity in cells or protects animals from venom, so the work has not produced a validated antivenom.
What the discussion said
Commenters spent less time celebrating scorpion antivenom than asking what Claude actually contributed and whether binder design moves the drug-development bottleneck. The charitable reading is that an LLM competently planned and ran a protein-design workflow: it selected established structure-design and scoring tools, generated a large candidate pool, and helped narrow it to lab-tested binders. Several readers see real value in that orchestration, especially for making sophisticated computational biology pipelines accessible to smaller teams. A prior snake-antivenom study with animal protection data also made the general technical direction look credible. The dominant caveat was that a designed binder is not a medicine. Clinical failure usually exposes weak target biology, toxicity, delivery, manufacturing, or efficacy assumptions made much earlier, while experimental validation remains slow and expensive. Readers stressed that this report showed binding results, not an antivenom proven in animals or people, and wanted peer review rather than a company announcement. They also pushed back on branding the achievement as Claude independently designing proteins: scientists specified targets, rules, tools, rankings, and diversity constraints, while specialized models did the molecular design work. The thread further noted a safety and access tension: Anthropic can pursue biologics under controlled partnerships while restricting comparable bio work by ordinary API users, and even a successful therapy may remain unaffordable without funding for trials and distribution.
Where opinion split
The sharp dispute is whether this is a meaningful AI drug-discovery advance or polished workflow automation. Supporters argue that reliably coordinating design tools and producing experimentally confirmed binders can raise the odds of downstream success and broaden access to computational biology. Skeptics answer that the model followed a scientist-authored recipe using existing specialist tools, while the decisive hurdles remain biological validation, clinical trials, manufacturing, and access.
Coverage & discussions
Hacker News · anthropic.com
Community discussions
Community Sentiment
Positives
- LLM-guided orchestration can turn a pile of specialized protein-design tools into a usable research workflow, lowering the expertise barrier for computational binder discovery.
- The team moved from hundreds of thousands of computational candidates to a small lab-tested set with multiple toxin binders, a concrete result rather than a purely simulated demo.
- Earlier AI-designed antivenom work reaching mouse-protection experiments suggests that computationally designed toxin binders can progress beyond attractive binding scores.
- Cheaper, more available antivenom would matter enormously for victims of severe scorpion stings, where current treatment can be financially out of reach.
Concerns
- A binder only matters after researchers identify the right biological target and show that binding it changes disease safely; better molecule generation does not solve that foundational uncertainty.
- The reported work stops far short of a medicine: binding assays do not establish protection, toxicity, dosing, manufacturability, or clinical effectiveness.
- Calling the result Claude-designed overstates the model’s role when researchers preselected the targets, toolchain, scoring formulas, and design constraints.
- The lack of peer-reviewed publication leaves readers unable to judge experimental rigor, novelty, and how much the LLM improved on conventional protein-design pipelines.
- Even a promising antivenom blueprint changes little if nobody finances human trials and broad production, particularly in a market already marked by extreme pricing.
