AI-driven wet lab uncovers DNA-editing enzyme system in bacteriophages
Anthropic's Bay Area biology lab, which uses its AI models for physical experiments, has identified a previously unknown enzyme system in bacteriophage DNA that can cut, copy, and paste genetic material, similar to CRISPR. The discovery was made largely by the Claude model, which spent 21 hours analyzing data with about 950 agents and 210 million tokens. Human researchers still conducted the physical experiments, and the finding awaits validation by the broader scientific community.
The new enzyme system, found in bacteriophage DNA, can perform genetic operations akin to CRISPR. The lab operates at low biosafety levels (BSL-1 and BSL-2) and avoids human-infecting pathogens. Human scientists handle all physical work, while Claude processed data using roughly 950 agents and 210 million tokens over 21 hours. The finding awaits peer review, and Stanford previously reported a somewhat similar system.
Anthropic's lab opened this spring, making the speed notable. Yet AI-driven biology is not unique; Stanford published on LLMs with CRISPR, UCSF designed enzymes with AI, and Google's AlphaFold launched in 2020. Amodei suggests Claude could eventually control lab equipment autonomously with safeguards, though that is not happening today.
This development could significantly accelerate biological discovery, potentially leading to new gene-editing tools and faster drug development. However, it may also heighten concerns about AI's role in dual-use research, as the same capabilities could be misused for bioterrorism. The reliance on AI for hypothesis generation may shift scientific workflows, but human validation remains essential. Society could benefit from improved healthcare, yet must carefully weigh the risks of autonomous AI in sensitive labs.