Anthropic’s Claude Yields 17 Scorpion Toxin Binders in Lab Tests

Anthropic’s AI model Claude produced 17 protein designs that met laboratory tests for binding to scorpion nerve toxins, according to results released on Oct. 20. This is an early finding in drug research. It does not show whether the designs can stop the venom from causing harm.

An internal AI model created 318,484 designs during a 64-hour effort and chose 180 for testing. Anthropic’s life sciences team added 31 more designs with Claude’s help, bringing the total to 211. Using Claude Science, the model did most of the work on its own while the team supervised it. The team guided the work with an initial prompt and four follow-up prompts.

The experiment focused on five nerve toxins: Cn2 from Mexico, Ts1 from Brazil, AaH I and AaH II from North Africa, and LqhIII from the Middle East. Seven designs met the binding requirements for Cn2, four for AaH I, four for AaH II, and two for LqhIII. No designs met the requirements for Ts1.

Adaptyv Bio and Twist Bioscience carried out the binding tests. Each of the 17 successful designs attached to just one toxin. Claude also created designs meant to attach to several toxins, but none met the requirements for binding to more than one. The tests checked whether the proteins attached to their targets, not whether they stopped the toxins from causing harm.

The two designs that bound most strongly to Cn2, called Cn2-B1 and Cn2-B2, had binding strengths in the single-digit nanomolar range. This is a way to measure how tightly they attached to the toxin. In tests where the proteins competed to bind, the two designs seemed to attach at or near the same spot as an antibody that has been shown to neutralize Cn2 in mice. This does not prove that either Claude-designed protein can neutralize Cn2.

The tests used toxins made in a laboratory, not toxins purified from actual venom. The strongest binders have not been tested in cells or animals. So, researchers do not yet know whether they can block the toxins’ effects or whether they are safe. Binding to a target in a lab is one step in checking whether a possible treatment might work, but these results do not amount to an antivenom.

This work provides another laboratory example for an AI drug discovery investment thesis that looks at whether designing possible drugs faster could create more work for testing laboratories and research equipment companies. In this experiment, outside laboratories tested a small group chosen from hundreds of thousands of computer-generated designs. Anthropic is making all 211 designs and their binding data available to other researchers.