Anthropic's Claude model designs 35,000 antibodies for food allergies

Summary

Anthropic's Claude model generated approximately 35,000 antibody designs in one month as part of a collaboration with Divergence Labs targeting IgE for food allergy treatment. Early-stage laboratory tests indicated that some candidates were more successful at dissociating from IgE than standard drugs.

Why it matters

This development represents a concrete step in the transition of large language models (LLMs) from text generation to industrial-scale structural molecular design. The process could reduce antibody engineering timelines from months to weeks. A six-partner drug pipeline involving Divergence, Generate Biomedicines, BigHat, Enveda, Bristol Myers Squibb, and Mithril signals a broad investment in AI-driven drug discovery. However, all results remain in the pre-clinical stage and have not been tested in humans.

Highlights

• Approximately 35,000 antibody designs in about 1 month; only 3 bound to IgE, 20 dissociated from IgE, and 3 outperformed the standard drug in dissociation tests
• Claude Code redesigned CAR T-cell genetic instructions, reducing length by approximately 70% while increasing cancer-recognizing receptor output tenfold
• In a 1,061-page document, pharmaceutical industry document review time was reduced by 89% (from 25+ hours to under 3 hours); all 11 issues detected by humans were caught, and 7 new issues were found

Critical take

The article reads more like a corporate press release than an independent journalistic piece. While every metric is framed as 'promising,' the critical caveat—that no candidate has been tested in even a single patient—is compressed into a single sentence near the end of the text. Although the figure of 35,000 designs appears impressive, it should not be overlooked that only 23 candidates showed relevant activity and that the two key functions (binding + dissociation) have not yet been combined in a single molecule. This indicates that a viable drug is still years away.