A Lab That Runs Its Own Molecular Biology Experiments

A team at Shanghai Jiao Tong University has built a laboratory system that runs molecular biology experiments on its own, and published the work in Nature, where it was featured on the journal's homepage.

The system, called AutoDNA, is an attempt to fix something specific about existing “self-driving labs”. Most of them are bolt-on: experts configure the instruments in advance, write the protocol in advance and set the optimisation rules in advance, and then AI is attached to one part of the loop — searching the literature, drafting a plan, analysing the results. That works when the workflow is simple and the instruments are few. Molecular biology is neither.

AutoDNA's answer is to let the model span the whole loop through co-design of model, experiment and instrument. A task is broken into stages — literature search, reagent selection, protocol design, code generation, hardware execution, result analysis, iterative refinement — and each is handled by a specialist agent. The agents coordinate as the experiment proceeds, so a failed run changes the next plan rather than just being recorded.

The second problem is hardware heterogeneity. Instruments differ by manufacturer and interface, and a model cannot be expected to learn each one. The team distilled what every device can do into what they call atomic services: a natural-language description plus a unified code interface, telling the model what the instrument does and how to call it. The abstraction hides the differences underneath and lets the model compose capabilities, generate cross-device control code, and check and repair it.

The third problem is physical. Pipetting a microlitre and positioning a plate to a micron is not a reasoning task, so the platform pairs high-precision automation with a scheduling layer that checks programs, tracks instrument state and allocates resources, keeping several experiments running within their timing and handling constraints. The control terminal runs OpenHarmony.

The group demonstrated the system on nucleic acid testing, enzymatic DNA synthesis and cancer detection. The significance is less the individual assays than the architecture: if a model can drive instruments it has never seen, through a description rather than bespoke integration, then the bottleneck for autonomous science moves from engineering to the quality of the questions asked.