Anthropic reports that an internal Claude research model helped design small proteins that bind to scorpion neurotoxins. Its October 20 account describes a computational campaign followed by laboratory binding tests, with substantial work still needed before any treatment claim.
From generated designs to measured binding
The model generated 318,484 designs and selected 180 for testing. Researchers added another 31 with Claude’s help, bringing the tested set to 211. Adaptyv Bio and Twist Bioscience performed the laboratory work.
Seventeen designs met the study’s binding criteria. The results were against individual toxins; none of the designs intended to address multiple toxins met the criteria for more than one. The tested subset and the initial generated collection are therefore different denominators and should not be used interchangeably.
Binding is not demonstrated protection
Anthropic explicitly says it has not confirmed that the designs neutralise scorpion venom in a cellular environment. The reported study did not test treatment effectiveness in living animals.
Attaching to a target is an early research result. It does not establish that a molecule blocks the toxin’s harmful function, is safe or can become an affordable antivenom. The announcement itself identifies later validation and practical development as remaining challenges.
What the release makes possible
Anthropic is releasing all 211 designs and their binding data for other researchers. The account also describes four human follow-up prompts and researcher supervision, so it should not be presented as an entirely unassisted therapeutic discovery.
Franklin’s interpretation is that the most useful contribution is a testable set of candidates and measurements. Follow-up work can examine whether binding translates into the intended function, and whether promising individual results survive further validation. The current evidence supports a research milestone, not a new medicine.
Read Anthropic’s study account for the design campaign, assay criteria, limitations and data release.