1 of 3 free pieces this week. Become a member to read every member piece. It's free with a work email or LinkedIn.
What's Happening
- Investment firm Citrini Research teamed up with Anthropic to test an unreleased version of the Claude AI model on drug development tasks.
- Anthropic dropped the restrictions on biosecurity and covered $50,000 in computing costs for the experiment.
- The AI generated 318,484 designs for protein binders targeting scorpion neurotoxins over a 24-hour period.
- Laboratory tests confirmed 17 of the 211 designs met the criteria for a binder, though none function as a drug.
Why It Matters
Anthropic wants to demonstrate AI capabilities in life sciences ahead of a planned public offering. The experiment produced viable protein binders but failed to yield a functional therapeutic. The company's decision to lift biosecurity guardrails for an investment firm demonstrates the tension between model safety and commercial promotion.
The Context
- Scorpion neurotoxins kill more than 3,250 people annually, and current antivenom production requires injecting venom into horses.
- Citrini Research published a piece of fiction detailing an AI-induced economic crisis that affected real-world markets.
- Anthropic operates a dedicated research environment called Claude Science to support biological and chemical engineering tasks.
What They're Saying
- Citrini Research acknowledged its team lacked biological expertise, stating that none of its members knew anything about AI as it related to biology.
- External experts Susana Vázquez Torres and Timothy P. Jenkins described the protein binder designs as a promising first step.
What to Watch
- Anthropic prepares for an IPO and continues to promote its models for life sciences applications.
- Researchers will test the two strongest protein binders to determine if they neutralize Cn2 in living organisms.
The Bottom Line
Anthropic and Citrini Research used an unreleased AI model to design protein binders for scorpion venom, achieving early scientific success but falling short of creating a usable drug.