Shanghai Jiao Tong University’s AutoDNA lab lets AI agents run DNA experiments

Researchers at Shanghai Jiao Tong University have built a lab where the scientists are mostly software. The system, called AutoDNA, uses a team of AI agents to design and carry out DNA experiments from a single typed instruction.

Humans still have a job. It just looks more like stocking shelves and fixing machines than running experiments.

The work is described in a paper published in Nature this week, dated November 4, 2026.

How AutoDNA works

Think of AutoDNA as a small research group in which every member is an AI agent with a specialty. One person types a request in plain language. The agents then split up the planning and execution between them.

On the physical side, robotic arms shuttle samples from one machine to the next. The system can run more than 20 types of laboratory equipment.

That matters because DNA work is rarely a single step. A single project can stretch into thousands of individual operations, including both synthesizing DNA and sequencing it.

The human role is narrow by design. People top up reagents, the chemical ingredients experiments consume, and keep the hardware maintained. Everything else, from experimental design to running the protocol, falls to the machines.

The research team was led by Haibo Chen, working alongside Mingyu Wu, Zhaoguo Wang and Jiabin Wang. They describe the approach as AI-native, built around multiple cooperating agents rather than one model trying to do everything.

What the tests showed

In testing, AutoDNA matched expert-level yields in enzymatic DNA synthesis. Put simply, when it built DNA strands using enzymes, it produced about as much usable material as skilled human researchers do.

The system also ran experimental protocols in parallel 3.2 times faster.

The most demanding test involved data storage. Researchers encoded digital information into DNA strands, and AutoDNA retrieved it exactly. That single task took 11,382 hardware steps and ran for roughly 203 hours.

The team also pointed AutoDNA at a medical problem. Working with a real-world clinical sample, the system correctly identified a KRAS G13D mutation, a genetic change considered clinically significant.

Background: the slow march of the self-driving lab

The idea of a lab that runs itself has been floating around research circles for years. Many automated setups handle one narrow workflow well and struggle the moment a project requires switching instruments or changing plans mid-stream. AutoDNA’s claim to novelty is that it handles end-to-end workflows across many machines with a single natural-language prompt at the front.

A preprint version of the work was posted on arXiv in July 2025, more than a year before the peer-reviewed Nature publication.

What this means

The diagnostic result deserves attention. Detecting the KRAS G13D mutation from a clinical sample hints at applications beyond pure research, including genetic testing workflows. Moving from an academic demonstration to a regulated clinical setting is a long road, though, and the paper does not claim AutoDNA is ready for hospital use.

The DNA data storage test points to a separate field. Encoding digital files in DNA has long attracted interest because of how densely genetic material can pack information. An automated system that can recover the data exactly, even over 11,382 steps, addresses one of the practical sticking points.

The research describes performance in the team’s own lab, and replication by other groups will be the real test. When an AI system designs its own experiments, researchers need ways to check its reasoning and catch errors before they consume days of machine time and expensive reagents. The fact that humans still restock and maintain the lab keeps people physically in the loop, but not necessarily intellectually.