Evo 2 is an AI model designed to read, interpret, and generate sequences of DNA, RNA, and proteins.Credit: Jean-Claude Revy/Science Photo Library
In 2008, researchers reported the first ever synthetic genome of a living organism, which was produced by chemically synthesizing the 580,000-nucleotide genome of the bacterium Mycoplasma genitalium1 . Follow-up work ‘rebooted’ such genomes in cells, creating what scientists claimed to be the first example of synthetic life2.
Now, researchers have used artificial intelligence to design whole genome sequences, including one inspired by that of M. genitalium. The AI model is trained on trillions of DNA letters from organisms across the tree of life.
Although impressive, the genome designs — included in a Nature paper, published on 4 March, describing the Evo2 DNA language model used to create them3 — are just one step towards creating AI-generated microbial life, say other researchers.
“It’s cool, but it’s not there yet,” says Nico Claassens, a synthetic biologist at Wageningen University in the Netherlands. The need to synthesize and test AI-generated genomes at scale is one major hurdle. Another is designing genomes that can direct all the essential functions of even the simplest life forms, let alone those of more-complex cells.
Yet scientists who have been working for more than a decade to design genomes from scratch say that this once-audacious goal now seems tangible. “These AI models are the ‘ChatGPT moment’ for synthetic genomics,” says genome engineer Patrick Yizhi Cai at the University of Manchester, UK. “You can start writing things that never existed in nature.”
Working genomes
Previous efforts to write genomes have mostly amounted to tinkering, the equivalent of editing a chapter of an existing book or removing all the commas, says Cai, who is part of a nearly finished project to rewrite the genome of baker’s yeast (Saccharomyces cerevisiae)4. Another effort recoded the genome of the bacterium Escherichia coli at nearly 20,000 locations, dispensing with 3 out of the 64 ‘codons’ that specify instructions for proteins5.
DNA language models, such as Evo2, raise the possibility of making synthetic life forms that differ more substantially from existing life. In 2025, Brian Hie, a computational biologist at the Arc Institute in Palo Alto, California, and his colleagues used Evo models to write genomes of bacteria-infecting viruses called phages6. When the researchers inserted the instructions into E. coli cells, 16 out of 285 designs produced functional viruses capable of killing bacteria.
Biggest-ever AI biology model writes DNA on demand
But the phage genomes were several thousand DNA letters long, encoding a handful of genes. This is vastly smaller and less complex than even small bacterial genomes, and most scientists see viruses as genetic parasites that lack many hallmarks of life.
In the latest Evo2 paper, a team led by Hie and Arc Institute bioengineer Patrick Hsu used the model to create the M. genitalium-inspired genome, as well as those of human mitochondria and a yeast chromosome.
Computer predictions suggested that nearly 70% of the genes in the M. genitalium-inspired sequences looked realistic. But if even one essential gene is missing or poorly modelled, the genome wouldn’t work inside a cell, says Claassens. “You cannot design life 70%. You can do that on a computer, but it will not be functional.”
Even if all essential genes were included, how those genes are ordered could also make or break a design, says Maciej Wiatrak, a machine-learning scientist at the University of Cambridge, UK, who has developed another AI tool, Bacformer, capable of generating bacterial genomes7. “Evaluating whether your genome looks correct and works correctly are two different beasts.”
