Claude Discovers CRISPR-Like Enzyme System in Phages
Anthropic said Claude discovered ART, a CRISPR-like enzyme system, in bacteriophage DNA. 950 Claude agents screened 200,000+ reverse transcriptases in 21 hours, shortlisting 20 candidates, with humans running verification in BSL-1/BSL-2 labs. ART's function remains unclear.
Key insights
Claude identifies ART, a CRISPR-like reverse-transcriptase system
Anthropic announced that Claude discovered a previously unnoticed enzyme system in bacteriophage DNA whose structural features resemble CRISPR, naming it ART — Array-associated Reverse Transcriptase. ART consists of a reverse transcriptase, a partner gene of unknown function, and a long array of equally spaced repeated DNA sequences. Experiments confirmed the array is transcribed into a set of distinct short RNAs, hinting at a programmable mechanism like CRISPR's. Systems combining these features have been found only rarely before, and all proved able to cut, copy, and paste DNA. Anthropic acknowledged ART's precise function is not yet clear; Amodei said its biotechnological utility and significance are undetermined, though it is work he would have been proud to do as a PhD student.
Discovery ran at scale: 950 agents, 21 hours, 210 million tokens
Researchers gave Claude only an initial prompt: find interesting new reverse transcriptases in massive DNA databases. About 950 Claude agents ran for 21 hours and consumed 210 million tokens, screening 3,500 candidates from more than 200,000 reverse transcriptases and narrowing to the 20 most worth examining, written up as human-readable reports. One agent spotted the repeat arrays while reading raw sequences, counted repeats, measured spacing, compared against known systems, and checked the literature before submitting for human review. The reverse transcriptase itself had been reported before, but Claude is credited as first to identify the adjacent array and partner protein. Genome mining of this kind typically takes human experts weeks to months.
Human scientists handled all experiments in a BSL-1/BSL-2 wet lab
The discovery came from Anthropic's life sciences team formed this spring, which built its own wet lab in the Bay Area restricted to BSL-1 and BSL-2 research and not touching pathogens that infect humans. All experiments were performed by human scientists; Claude proposed the verification experiments and the team carried them out. The team's daily tools are the publicly released Claude Science and Claude Code, meaning other labs can in principle replicate the 'AI generates hypotheses in bulk, humans validate by experiment' workflow. Anthropic is soliciting research topics from external scientists and plans to grow the life sciences team.
Amodei argues biology AI follows the same exponential as math
Dario Amodei rejected readings of the discovery as a one-off, citing a repeated pattern in which AI performance in new intellectual domains goes from weak to superhuman within a few years: in 2023 models struggled with average high-school math, in 2024 they did well on elite competition math, in 2025 they solved minor open problems, in early 2026 more significant ones, and by late 2026 they began solving the top few open problems in all of mathematics. He believes AI for biology is on a similar exponential trend, with experimentation the main difference: humans collaborate with AI to validate results in weeks, and Claude may eventually safely control lab equipment with safeguards — though not today.
Accelerating discovery increases pipeline throughput, not trial latency
Amodei mapped the biomedical pipeline — fundamental discovery, translational research, drug discovery, clinical trials, and delivery — and argued that accelerating only the first stage can speed and broaden everything downstream by identifying new drug targets, new therapeutic modalities, and more precise measurement while shortening the experimental loop. This does not by itself reduce clinical trial times, but it could greatly increase the number of promising candidates entering the pipeline: a throughput gain even though latency stays. He tied this to his Machines of Loving Grace claim that AI could help 'cure most diseases in 5-10 years' if applied to every pipeline stage.
Discovery joins a lineage, with independent Stanford parallel work
Amodei placed the finding in a decades-long lineage running from CRISPR through the bridge recombinase and VIPR in recent years, amid heightened interest in reverse-transcriptase-based systems. A Stanford team working independently recently described a novel RT system with an associated non-coding array similar in some ways to Claude's find, though the two systems evolved independently. CRISPR pioneer Feng Zhang of MIT reviewed the preprint and called it a good example of AI agents participating in biological discovery, saying such arrays deserve deeper study. Anthropic's position is that the field is at the very beginning of finding such systems and developing them into powerful biotechnology tools.