Anthropic’s new life-sciences laboratory has reported an early result from a large AI-assisted genome-mining campaign. Claude agents identified an unusual arrangement around a reverse transcriptase in bacteriophages: the enzyme gene, an adjacent partner gene and a long array of evenly spaced DNA repeats. The team calls the system array-associated reverse transcriptases, or ART.[1][2]
The result is easy to compress into “AI discovered a new enzyme system.” The more instructive story is the sequence of evidence. A model searched and prioritised; scientists reviewed the report; laboratory experiments checked whether the genomic pattern corresponded to a real biological system; and the team is still trying to establish its function.
The reported search funnel
| Measure | Value | Context |
|---|---|---|
| Parallel agents | Roughly 950 | Agents searched and investigated different reverse-transcriptase families.Source 1 |
| Search time | 21 hours | Elapsed campaign time, not 21 hours of total compute.Source 1 |
| Token use | 210 million | About 221,000 tokens per agent on average if divided evenly; actual use need not have been uniform.Source 1 |
| Reverse transcriptases gathered | More than 200,000 | The broad sequence collection searched by the agents.Source 1 |
| Candidate systems | 3,500 | New candidates identified before further narrowing.Source 1 |
| Detailed reports | 20 | The most compelling candidates selected for human-readable analysis.Source 1 |
What the AI actually contributed
The underlying reverse transcriptase had appeared in previous research. Anthropic’s claim is narrower and more interesting: Claude noticed the defining combination of the enzyme, non-coding repeat array and accessory protein, then compared the arrangement with known systems and searched for prior descriptions. It produced a report for human review rather than a final biological conclusion.[1][2]
This is a plausible division of labour for discovery. Sequence databases are too large for a scientist to inspect manually, while a general model can create many hypotheses and narrative reports. The human bottleneck shifts from finding every anomaly to judging which proposals are coherent, novel, experimentally reachable and worth scarce laboratory time.
Four stages that should not be collapsed
| Stage | Evidence produced | What it does not establish |
|---|---|---|
| Database search | A recurrent genomic pattern and candidate family | Biological function |
| AI hypothesis | A proposed system structure and analogy with known mechanisms | That the analogy is correct |
| Human review | A candidate judged worth experimental effort | Experimental validation |
| Laboratory work | The repeat array is expressed as distinct short RNAs | The primary role or programmability of ART |
The CRISPR comparison needs precision
ART’s repeat layout resembles a CRISPR array, and the published team says the combination of features has appeared together in only a small number of other systems that can perform programmable DNA operations. That makes ART scientifically intriguing. It does not mean ART has already been shown to cut, copy or paste DNA in a useful programmable way. Anthropic explicitly says its primary function remains unknown.[1]
How to judge the next AI-assisted discovery
The most useful scorecard should record the complete funnel. How many hypotheses were generated? How many survived expert review? Which experiments were run? How often did the predicted mechanism fail? Were negative results preserved? Can another laboratory reproduce the observations? A high number of candidate reports is valuable only if the system improves the yield or speed of validated findings.
Anthropic’s Claude Science workbench emphasises auditable artifacts and access to scientific tools. That matters because discovery claims need more than a persuasive final answer: they need the database queries, candidate lists, code, assumptions and experimental record that connect the initial search to the result.[3]
The ART result is therefore best understood as evidence for a new research loop. AI can scan a space too large for manual inspection and propose anomalies worth attention. Scientists still establish which anomalies survive contact with the laboratory. The opportunity lies in making that loop faster, more transparent and more productive—not in pretending that a model-generated hypothesis is already a discovery.
Sources & further reading
- Claude discovers a novel enzyme system with CRISPR-like repeats Anthropic · 23 Sep 2026 · Accessed 9 Oct 2026
- Array-associated reverse transcriptases technical report Anthropic · 23 Sep 2026 · Accessed 9 Oct 2026
- Claude Science, an AI workbench for scientists, is now available Anthropic · 30 Jun 2026 · Accessed 9 Oct 2026