Scientists at Stanford University and the Arc Institute, a nonprofit research lab in Palo Alto, say they have produced the first complete genome designed entirely by artificial intelligence. The organism is a bacteriophage, a virus that infects and kills bacteria, and the advance paper has drawn wide attention as a step towards machine-designed life, as reported by MIT Technology Review.
The team worked with phiX174, a small bacteriophage with just 11 genes and roughly 5,000 letters of DNA. They trained Evo, a large language model built for biological sequences, on the genomes of about two million other bacteriophages. The model was fed no text at all, only DNA.
To test whether machine-written genomes could work, the researchers chemically printed 302 designs and mixed them with E. coli bacteria. One evening they came back to the lab and saw dead patches, clearings in the bacterial lawn where viruses had multiplied and burst their hosts. Microscope images confirmed tiny viral particles. Sixteen of the 302 computer designs replicated and killed the bacteria.
Jef Boeke of NYU Langone Health, who has worked on synthetic genomes for decades, called the advance copy an impressive start. He said the model surprised him by proposing new gene sequences, shortened genes and reordered arrangements that human designers had not tried.
The work sits in a longer line of synthetic biology. In 2008 a team led by J. Craig Venter built the first bacterium with a lab-printed genome after testing many candidate designs by hand, work he describes as manual AI. Venter now sees machine learning as a faster version of the same trial and error. Interest in the field is rising fast: protein structure prediction won a Nobel prize in 2024, and the Boston company Lila this week raised $235m for automated, AI-run laboratories.
Practical uses are closest for medicine. Doctors have long used phage therapy to attack stubborn bacterial infections, a problem that keeps growing as antibiotic resistance in bacteria spreads, and similar tests target black rot in cabbage. Samuel King, the student who led the project, said AI could also design better viral vectors for gene therapy, an idea related to work on DNA-built nanoscale tools.
The researchers say they deliberately never trained the model on viruses that infect humans. Even so, they acknowledge the method could in principle be pointed at human pathogens, a dual-use risk. Jason Kelly, chief executive of cell engineering firm Ginkgo Bioworks, believes the next leap needs automated labs that propose, build and test genomes in a loop, feeding results back to the model. He called it a breakthrough of national importance, since cells underpin every living thing.