Live·Open questions in longevity research
All news
Scientific ComputingScience Research

Following Claude’s plan, Ginkgo’s robotic lab tested four enzymes from a family discovered by AI agents in 13 days: two copy RNA into DNA

18 November 2026· 261118012

Following Claude’s plan, Ginkgo’s robotic lab tested four enzymes from a family discovered by AI agents in 13 days: two copy RNA into DNA

On November 17, Ginkgo reported that Claude Opus 5 had designed a set of experiments that humans reviewed. The first data arrived 13 days after the plan was drawn up, and the laboratory services cost about 14 thousand dollars. The enzymes SA1 and LY01 copy RNA into DNA; the other two, MarsHill and PB50, produced weak signals or signals at background levels.

In September, Anthropic, the developer of Claude, announced that its AI agents had discovered the ART system in the genomes of phages (viruses that infect bacteria). The system contains a gene for reverse transcriptase (an enzyme that copies RNA into DNA) and a repeat array: a short sequence appears repeatedly, with different spacers between the repeats, as in CRISPR. Individual short RNAs are transcribed from the array. The authors of the preprint proposed that the system works like retrons, bacterial defense systems against phages. In retrons, an enzyme copies a particular RNA into DNA, which then contributes to defense. The authors also wrote: “We have not shown that the reverse transcriptase is active or that the RNA units serve as its substrates.”

Ginkgo set out to test this. The company sells experiments through a cloud laboratory, where customers assemble workflows from a catalog of steps and robots carry them out. According to the report, Ginkgo built a software wrapper around Claude Opus 5 that broke the preprint down into claims, identified gaps, and attempted to refute each point it identified. The points that survived became an experimental plan: a sequence of verified steps that could be completed in 13 days. Four of the eight ART systems in public databases were selected: SA1, MarsHill, LY01, and PB50.

Claude specified the DNA sequences encoding the proteins and the assay parameters, then submitted the orders. Humans reviewed the plan. They prioritized cell-free synthesis (producing proteins in a test tube using bacterial cell contents) and catalytically “dead” control versions of the enzymes. These controls have substitutions at two amino acids in the active site and are expected to be unable to copy RNA. When capillary electrophoresis failed to detect most of the full-length proteins, the researchers added a luminescence-based protein assay. “Execution still relied on close human oversight,” the report’s authors write. Anthropic CEO Dario Amodei wrote on September 23 that Claude might someday be able to conduct experiments safely by controlling equipment autonomously, but Anthropic does not do this today.

Each phage enzyme was supplied with synthetic RNA from its own array, and activity was measured by quantitative PCR: the earlier the signal from the DNA copy rises, the more copies the enzyme has produced. No primer, from which copying usually begins, was added, so the enzyme had to initiate the copy itself. The researchers tested for a signal at three regions of the array. Two span repeat-spacer junctions, where a signal requires a continuous DNA copy across the junction. SA1 and LY01 produced signals above background at all three regions. Signals remained at background levels for the “dead” versions and for versions lacking the unusually long N-terminal region of the protein (about 180 amino acids, compared with the usual 50 or fewer), although removing this region made protein production more reliable.

The report’s authors write: “We do not yet have an explanation for why two candidates are active and two are not.” Electrophoresis showed that tagged versions of SA1 and MarsHill yielded nearly equal amounts of protein, but the SA1 signal appeared 12,7 PCR cycles earlier than that of its “dead” version, whereas the MarsHill signal appeared 1,6 cycles earlier. The authors plan to sequence the synthesized DNA and test in living cells whether transcription of the array protects against phages.

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-agents#robotic-lab#reverse-transcriptase#phage-defense#rna-to-dna#art-system