AI agents run DNA experiments at Shanghai’s AutoDNA laboratory, while humans maintain the equipment

AI agents run DNA experiments at Shanghai’s AutoDNA laboratory, while humans maintain the equipment

A paper published in Nature on November 4 describes a laboratory at Shanghai Jiao Tong University. Seven agents based on the Gemini 2.5 Pro language model develop experimental protocols, write code for 21 types of instruments and execute it.

James Zou of Stanford University writes in a commentary on the paper:

“An AI system can design a promising biologically active molecule in minutes, but testing it can take weeks or months.”

Co-author Haibo Chen says in a Nature news article that specialists still configure the instruments in many automated laboratories, with AI usually added to just one part of the process. His team built AutoDNA around AI from the outset. A planning agent receives a task in plain language and calls on other agents, while a dedicated agent converts instrument manuals into descriptions the other agents can use. The pipette manual lists 19 commands; the description reduces them to two actions: transfer and mix.

One experiment involved enzymatic DNA synthesis: an enzyme extends DNA strands one nucleotide at a time while the strands are attached to magnetic beads. The literature agent found a benchmark: a 2022 study using this enzyme reported an average step yield (the proportion of strands that received the intended nucleotide) of 98,7%. The task was to achieve that yield in the shortest possible time and stop if the yield fell below 97,7%. The starting protocol produced a yield of 94,07% in 6 hours 48 minutes. The planning agent reversed one change proposed by the hypothesis agent because it reduced the yield. Two other proposals were dropped because the reagent agent could not find the required solutions in stock. The hypothesis that clumped beads were blocking the enzyme’s access to the strands led to the addition of a detergent to the wash. This change, together with higher reagent and enzyme concentrations, brought the yield to 97,66%. The agents then began reducing the time: first to 6 hours with a yield of 98,0%, then to 5 hours 20 minutes with a yield of 97,76%. An attempt to shorten the washes reduced the yield to 96,86%, below the threshold, and the planning agent stopped.

Reviewers asked the authors to compare the agents with Bayesian optimization, an algorithm trained on data from published papers to select experimental conditions. The agents brought the step yield to 98,0% over nine rounds of experiments; the algorithm reached 92,6% in the same number of rounds.

The scheduler that executes the agents’ code combines identical steps from different protocols into a single instrument run. Running four variants of the synthesis protocol sequentially took an average of 1434 minutes. Combining their steps reduced this to 454, making execution 3,2 times faster.

In the longest experiment, the agents devised an encoding scheme and stored the English sentence “Intelligence is the ability to adapt to change” in DNA. A sequencer, an instrument that reads DNA, read the strands, and the sentence was recovered without errors. The experiment involved 11 382 operations and lasted 203,5 hours.

In an experiment with a colon tumour sample, the system had no prior knowledge of its sequence. The laboratory sequenced the sample’s DNA and found the G13D mutation in the KRAS gene, which had also been identified by clinical testing. The agents designed primers, short DNA strands that initiate DNA synthesis during PCR, and synthesized them using the protocol developed during synthesis optimization. Testing on the first attempt showed that the primers distinguished mutant DNA from normal DNA.

When reagents or consumables run out, the scheduler pauses the experiment and calls an operator. According to the supplementary materials, during the experiment that stored the sentence, the operator replenished one solution 79 times and pipette tips 108 times. A reviewer considered the phrase “without human intervention” too strong, and the authors clarified that human intervention was limited to equipment maintenance.