A Shanghai lab built for AI agents runs DNA work, with people on call

AutoDNA's AI agents ran DNA synthesis and sequencing cycles in about 200 hours, but people still restocked supplies and fixed 6 of 14 hardware errors.

Researchers at Shanghai Jiao Tong University have built AutoDNA, a biology lab designed around AI agents from the start instead of adapting a human lab. Several large-language-model agents share the work, including a coordinator that plans and others that write code to drive the hardware. Robotic arms operate more than 20 types of common lab equipment and can correct small hiccups themselves.

A scientist gives the system a task in plain language. The agents devise the procedure, generate control code, and carry out DNA synthesis, amplification and sequencing. A full DNA read-write cycle took about 200 hours, with precision comparable to expert scientists. In one test, the system tuned PCR reaction settings over several rounds and reached a yield near 98 percent, similar to human experts. In another, it flagged a candidate KRAS mutation in a colorectal cancer sample, then designed and synthesized primers to confirm it. In a message-encoding run, it performed 11,382 operations over roughly eight days, and all 78 synthesized DNA strands were made and decoded despite a few missing DNA letters.

Humans were still in the loop. People restocked reagents and disposable labware and emptied the trash, and the system paused and alerted a technician when supplies ran short. Of 14 hardware errors logged, 6 needed a person to fix. It cannot yet handle delicate experiments with living cells or animals.

The abundance case is about labor and time. If routine bench work can be delegated to agents, experiments get cheaper and faster, and scientists can spend more effort on results and hypotheses. That matters for health research, where steps like confirming a cancer mutation are repetitive.

The caveats are real. Outside experts were not involved, and one, Reshma Shetty of Ginkgo Bioworks, called it a teaser of future possibilities. The work is not ready for prime time. Unfamiliar, expensive equipment is a concern, with a digital twin suggested as a fix. Biosecurity is another: without safeguards, self-driving labs could synthesize DNA for dangerous proteins, and screening software, access control and an authorization ledger are proposed mitigations.

Source: Singularity Hub