Claude found a new enzyme system, and gene-editing stocks dropped
950 agents. 21 hours. 210 million tokens. One new enzyme system nobody had described before.
Anthropic pointed Claude at a giant database of DNA sequences with essentially one instruction and then got out of the way. Their own description of their involvement is blunt: they wrote the prompt and they did the lab work. Everything in between was agents. After 21 hours of roughly 950 of them combing sequences, one agent stopped on an unusual reverse transcriptase and noticed something sitting next to it β a tandem repeat array, evenly spaced, the structural signature of CRISPR. The agent's own note reads like a person double-taking: that's a CRISPR-like repeat array?!
What it found is now called ART, array-associated reverse transcriptases. Three parts: the reverse transcriptase itself, a partner gene of unknown function sitting beside it, and the long array of evenly spaced repeats. It shows up mostly in bacteriophages, the viruses that infect bacteria. The reverse transcriptase family was not new. The system was. Previous work had the enzyme and missed both the array and the accessory protein, which are exactly the two things that make it look programmable.
Anthropic's in-house lab then expressed the protein in laboratory strains and found the array is transcribed as a set of distinct short RNAs. That is the detail that matters. Short guide-like RNAs plus an enzyme is the shape of every programmable biology tool anyone has built. Feng Zhang, who has as much standing to judge this as anyone alive, reviewed the findings and endorsed the work. The preprint is titled "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays" and has not been peer reviewed.
The market reaction told its own story: gene-editing stocks slid on the announcement. That is investors pricing in the possibility that the discovery bottleneck in this field was never sequencing capacity or lab throughput but the number of trained eyes available to look at everything. 950 agents for 21 hours is a quantity of attention no lab can buy.
The honest caveat is the one Anthropic itself draws: a preprint, not peer reviewed, with the biological function still unknown. Nobody has shown ART does anything useful yet. But the process claim is separable from the biology claim and it is the bigger one. An agent swarm read a database, noticed a structure, and a wet lab confirmed the structure exists. Whatever ART turns out to do, that loop ran.
https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
← Back to all articles
Anthropic pointed Claude at a giant database of DNA sequences with essentially one instruction and then got out of the way. Their own description of their involvement is blunt: they wrote the prompt and they did the lab work. Everything in between was agents. After 21 hours of roughly 950 of them combing sequences, one agent stopped on an unusual reverse transcriptase and noticed something sitting next to it β a tandem repeat array, evenly spaced, the structural signature of CRISPR. The agent's own note reads like a person double-taking: that's a CRISPR-like repeat array?!
What it found is now called ART, array-associated reverse transcriptases. Three parts: the reverse transcriptase itself, a partner gene of unknown function sitting beside it, and the long array of evenly spaced repeats. It shows up mostly in bacteriophages, the viruses that infect bacteria. The reverse transcriptase family was not new. The system was. Previous work had the enzyme and missed both the array and the accessory protein, which are exactly the two things that make it look programmable.
Anthropic's in-house lab then expressed the protein in laboratory strains and found the array is transcribed as a set of distinct short RNAs. That is the detail that matters. Short guide-like RNAs plus an enzyme is the shape of every programmable biology tool anyone has built. Feng Zhang, who has as much standing to judge this as anyone alive, reviewed the findings and endorsed the work. The preprint is titled "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays" and has not been peer reviewed.
The market reaction told its own story: gene-editing stocks slid on the announcement. That is investors pricing in the possibility that the discovery bottleneck in this field was never sequencing capacity or lab throughput but the number of trained eyes available to look at everything. 950 agents for 21 hours is a quantity of attention no lab can buy.
The honest caveat is the one Anthropic itself draws: a preprint, not peer reviewed, with the biological function still unknown. Nobody has shown ART does anything useful yet. But the process claim is separable from the biology claim and it is the bigger one. An agent swarm read a database, noticed a structure, and a wet lab confirmed the structure exists. Whatever ART turns out to do, that loop ran.
https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
Comments