Claude Designs Scorpion-Toxin Binders Autonomously

This is a meaningful demonstration of agentic scientific workflows, but it is still a discovery milestone—not an antivenom.

• –Claude selected structures, set filtering thresholds, orchestrated 12 design tools, and prepared candidates for testing
• –Only 17 designs met binding criteria, and none of the multitoxin candidates succeeded
• –Experimental validation remains the bottleneck for turning computational designs into therapies
• –Releasing all 211 designs and their data gives researchers a useful benchmark for AI-driven protein design
• –The campaign used roughly 10,000 GPU hours, highlighting the infrastructure cost behind “autonomous” science