← Back to Home

Letting AI Complete the Experiments: AutoDNA Connects DNA Design, Synthesis and Readout

A Shanghai Jiao Tong University team used AI agents to coordinate more than 20 types of instruments, demonstrating DNA information storage and retrieval and cancer-related mutation detection. Experimental automation has taken a step forward, from proposing protocols to handling reagents, but replenishing consumables and repairs still depend on people.

By SURL BioNews

A written experimental protocol is often still separated from reliable biological results by pipetting, reactions, sequencing and repeated adjustments. AutoDNA, developed by a team at Shanghai Jiao Tong University, attempts to bridge that gap: artificial intelligence agents plan DNA experiments, direct instruments to carry out the operations and revise the next steps based on the results. The significance of this research lies in extending AI's capabilities from computational reasoning to the laboratory bench, where real reagents and samples must be handled.

The study was published in Nature on November 4 and was also covered in the journal's news and commentary that day. According to Shanghai Jiao Tong University's Zhangjiang Institute for Advanced Study, AutoDNA was developed jointly by computer science and chemistry teams, drawing on the DNA Storage Research Center. The platform integrates more than 20 types of instruments, covering operations including nucleic acid synthesis, amplification and sequencing, enabling the system to turn a user's task into an executable experimental workflow.

A key element is collaboration among multiple specialized AI agents, which divide tasks including literature searches, reagent selection, workflow design, control code writing and results analysis. The functions of different instruments are organized into a unified software interface that the agents can combine and invoke. The university said the platform has micrometer-scale positioning and microliter-scale liquid-handling capabilities. An execution system checks code, tracks instrument status and allocates resources so that operations across devices follow the timing and sequence required by the experiment.

DNA information storage provided a comprehensive test. This technology first encodes digital content into DNA base sequences, writes it through synthesis and ultimately retrieves it through sequencing and decoding. According to results released by the university, AutoDNA linked encoding, synthesis, storage, sequencing and decoding, performing 11,382 hardware operations over the full read-write cycle and accurately recovering the original information. This result demonstrates whether a long sequence of steps can be connected successfully; this demonstration alone does not establish that DNA storage is commercially competitive.

Another demonstration was more closely related to biomedical applications: the system identified and validated the cancer-related KRAS G13D mutation. The paper also reported that AutoDNA achieved yields comparable to those of an expert-optimized protocol in enzyme-mediated DNA synthesis experiments. The former demonstrated workflow integration from analysis to validation, while the latter tested whether the system could adjust reaction conditions to a useful level. Both come closer to the capabilities laboratories actually need than simply generating a seemingly plausible set of operating instructions.

However, a research demonstration detecting a specific mutation is insufficient to establish clinical diagnostic accuracy, nor can it directly support the conclusion that the system would be equally reliable with different samples and unknown problems. AutoDNA still requires personnel to replenish reagents and consumables and maintain hardware. Its autonomy rests on the equipment and operational capabilities already integrated into the platform. Whether it can be extended successfully to experimental requirements beyond that configuration still requires separate validation.

AutoDNA gives the discussion of autonomous laboratories a more concrete benchmark: AI must do more than propose the next step—it must also ensure that reagents enter the correct containers at the correct time and feed the results back into the next round of decisions. For it to become a tool that other laboratories can adopt, it remains to be seen whether different teams can reproduce the results and what labor costs deployment, maintenance and handling exceptions will entail.

References

  1. Nature
  2. Shanghai Jiao Tong University, Zhang Jiang Institute for Advanced Study
  3. Nature