An unreleased version of Anthropic’s Claude model generated 320,000 protein candidates in an attempt to design a new scorpion antivenom, but the experiment failed to produce a viable drug. Citrini Research, the investment firm behind the project, admitted its team lacked biological expertise, yet they collaborated with Anthropic to test AI’s limits in life sciences just weeks before Anthropic’s anticipated IPO. The process involved significant costs, including a compute bill just shy of $50,000 and testing costs of approximately $150,000, but experts confirm the results remain far from a marketable treatment.

The Experiment Mechanics

Citrini set the AI loose with a 35-page prompt and 300 subagents, a task Anthropic facilitated by removing biosecurity guardrails and covering the compute cost. The model spent 64 hours whittling down candidates, eventually providing 211 designs for external testing at Adaptyv Bio and Twist Biosciences. Seventeen of these proteins successfully bound to four of the five targeted scorpion toxins, a technical win that Citrini framed as a success for AI-enabled discovery.

Scientific Reality Check

Despite the binding results, experts from the Technical University of Denmark and Khalifa University argue the excitement is premature. Andreas Hougaard Laustsen-Kiel noted that while discovering binders is "easy-ish," translating them into medicine is the hard part, a challenge involving safety, stability, and biological behavior that remains untested. Professor Andreas Bender further criticized the novelty, pointing out that high-throughput screening already finds hits at manageable costs without AI, and that protein design is largely a solved problem compared to the bottleneck of clinical translation.

Investment Angle Over Science

Citrini, known for a viral piece of fiction about AI economic catastrophe, positioned the experiment as an investment thesis rather than a biotech breakthrough. The firm accused critics of nitpicking and clarified its role as an investment research entity, not a startup, while highlighting public companies like Charles River Laboratories in its paid subscriber report. Anthropic, which did not respond to requests for comment, benefited from the publicity by showcasing its capabilities in the life sciences sector, even if the specific drug candidate remains theoretical.

Key Takeaways

  • Claude generated 320,000 candidates, reduced to 211 designs, with 17 binding to target toxins.
  • Compute costs were just shy of $50,000, with additional testing costs around $150,000.
  • Experts warn that binding to a toxin does not equal neutralizing it or ensuring safety.
  • Citrini is an investment firm, not a biotech developer, and did not pursue further clinical steps.

The Bottom Line

This experiment proves AI can handle the initial filtering of molecular candidates, but it highlights that drug discovery remains a biology problem, not a compute problem. Without clinical validation, binding affinity is just a promising starting point, not a product.