Anthropic releases 211 Claude-designed proteins; 17 bind scorpion toxins

Anthropic released 211 protein designs and their data on October 20, 2026, after a Claude-led campaign produced 17 proteins that met laboratory binding criteria against scorpion toxins. The result connects AI-generated designs to physical experiments. But it establishes attachment to toxins—not that the proteins can stop venom from harming cells or animals.

An autonomous run with human steering

The project began with a question from investment research firm Citrini Research: could motivated nonexperts use Claude to tackle a genuine health problem? Anthropic’s life sciences team chose scorpion neurotoxins as the target, hoping to generate starting points for a future antivenom.

An internal research version of Claude worked for 64 hours, using an initial prompt and four follow-ups. The prompt specified targets, available design and folding tools, and rules for novelty, scoring and filtering. Claude chose toxin structures, selected sites for its proteins to attach, and set quality thresholds.

Claude used 12 tools to generate 318,484 designs and predicted the combined structures of 553,052 design-toxin pairs. The campaign consumed roughly 10,000 GPU hours, a measure of computing use rather than elapsed time. Claude selected 180 candidates; Anthropic’s team added another 31 with its help.

Human intervention affected the selections. Most of Claude’s final candidates, including its strongest binder, were chosen only after the third follow-up. The follow-ups mainly pushed it toward proteins that could bind several toxins. Under supervision, Claude also analyzed the experimental data, generated figures and drafted the final report.

Two labs, four successful toxin targets

Adaptyv Bio and Twist Bioscience produced the proteins using different systems: Adaptyv used protein-making machinery extracted from bacteria, while Twist used mammalian cells. Both measured binding with surface plasmon resonance, which tracks how quickly molecules attach and separate.

To qualify, a design had to show binding in every repeat at both labs, with a dissociation constant of one micromolar or lower. That number describes binding strength; lower means tighter attachment. It also had to avoid binding an unrelated human protein used to detect nonspecific stickiness.

• Mexico’s Cn2 toxin: seven qualifying binders.
• North Africa’s AaH I and AaH II toxins: four binders each.
• The Middle East’s LqhIII toxin: two binders.
• Brazil’s Ts1 toxin: no designs met the binding criteria.

Every qualifying design bound just one toxin. None met the criteria for binding multiple toxins, a goal intended to reduce the number of ingredients in a broadly useful antivenom. Among the 60 designs aimed at individual toxins, 11 qualified—an 18% rate Anthropic describes as consistent with typical protein-design campaigns.

The challenge is to use these extraordinary tools to develop molecules that can become real drugs, ones that are safe, affordable, and widely accessible.

Susana Vázquez Torres, postdoctoral researcher at the Spanish National Cancer Research Centre, in Anthropic’s research account

A promising attachment site, not proven protection

The strongest designs, Cn2-B1 and Cn2-B2, bound Cn2 with single-digit nanomolar affinities at both labs. Competition experiments suggested they attach on or near the site used by LR, an antibody reported to neutralize Cn2 in mice. That overlap offers a reason to investigate whether the new proteins can block the toxin.

The intended mechanism is straightforward: dangerous scorpion neurotoxins attach to sodium channels, proteins that help nerve cells fire. A binder covering the toxin’s channel-contacting surface could prevent that attachment. But these experiments used toxins grown in the lab from their genes, and did not test neutralization in cells or protection in animals.

Existing antivenoms generally rely on antibodies purified from immunized animals, a production process that can take months. Anthropic’s proposed alternative starts with small proteins designed from scratch. The company is making all 211 designs and their data available for further research, rather than claiming to have produced an antivenom.