Claude discovers CRISPR-like ART enzyme system in phages
Anthropic announced on September 23, 2026 that Claude independently identified a new enzyme system in phage DNA named ART (Array-associated Reverse Transcriptase). It features a reverse transcriptase, an unknown partner gene, and evenly spaced repeated DNA arrays similar to CRISPR. The system was found by 950 Claude agents running for 21 hours, analyzing over 200,000 reverse transcriptases to produce 20 candidate reports.
Key insights
Claude's autonomous discovery process
Anthropic gave a single initial prompt to search for novel reverse transcriptases in large DNA databases. Roughly 950 Claude agents ran for 21 hours consuming 210 million tokens. They filtered over 200,000 candidates down to 3,500, then to the top 20, written as readable reports. One agent recognized the repeated array while reading raw sequences, measured spacing, compared with known systems, and confirmed the novelty before submitting for human verification.
ART structure and potential as a programmable gene editing system
ART consists of three components: a reverse transcriptase, a partner gene of unknown function, and long arrays of evenly spaced repeated DNA sequences. Experiments confirmed the arrays are transcribed into distinct short RNAs, suggestive of a programmable mechanism similar to CRISPR. Only a handful of systems with all these features are known, and all can edit, copy, or paste DNA. ART's exact function remains unknown.
Comparison to previous genome mining efforts and timeline
Human experts typically take weeks to months for this kind of genome mining. Claude completed the work in 21 hours of agent runtime. The reverse transcriptase enzyme itself had been previously reported, but Claude was the first to identify the adjacent array and partner protein. CRISPR pioneer Zhang Feng reviewed the preprint and called it 'a good example of AI agents participating in biological discovery.'
Broader implications for AI-driven biology and drug discovery
Anthropic CEO Dario Amodei noted this is part of a pattern where AI performance in intellectual domains goes from weak to superhuman within a few years, drawing a parallel to mathematics where AI has gone from competition-level to solving top open problems. He argued that biology requires experimentation, not just theory, but that human scientists can collaborate with AI to validate key results in weeks, and that even accelerating the first stage of fundamental biology discovery could speed up the entire biomedical pipeline.
Anthropic's wet lab and open collaboration approach
Anthropic's new life sciences team, established in spring 2026, operates its own wet lab in the Bay Area at BSL-1 and BSL-2 levels. All experiments are conducted by human scientists. The tools used—Claude Science and Claude Code—are publicly available, meaning other labs can replicate the AI hypothesis generation and human validation workflow. Anthropic is inviting external scientists to propose collaborative research projects.