Large genome models are beginning to show that AI can do more than suggest new proteins. In a recent test, researchers based at Stanford University used models trained on DNA to generate complete genomes for viruses that infect bacteria.
The work focused on ΦX174, a small virus that infects E. coli. The results were limited and carefully filtered, but they show why AI genome design is becoming a subject for both scientific interest and biosecurity caution.
Why DNA models are different from protein design
Much AI work in biology has centered on proteins because proteins carry out many of life’s practical tasks. They catalyze chemistry, help structure cells, and can provide new or useful functions when designed well.
DNA is less direct. The genetic code sits between a DNA sequence and the proteins it may encode, so it was not obvious what a model trained on DNA alone could reliably produce.
Even so, large genome models have already shown that they can output DNA sequences able to encode functional proteins in bacteria. They can also mimic gene structures found in complex cells.
The challenge is that DNA has both strict and flexible regions. Some positions matter intensely, while other stretches can vary without obvious effect. A useful model has to infer biological context from sequence alone, including context that human researchers may not fully understand.
How Evo 1 and Evo 2 were used
The models in this work were called Evo 1 and Evo 2. As a precaution during training, the researchers did not include sequences from viruses that target complex cells, because even unclear outputs could create concern if they had dangerous properties.
Instead, the team turned to bacteriophages, viruses that attack bacteria and are unable to infect humans. That made it possible to ask a sharper question: could a model generate not just a gene-like sequence, but an entire viral genome?
The chosen test case was ΦX174, a member of Microviridae. It infects E. coli and is small enough to analyze in detail: 11 genes across about 5,400 bases. All 11 genes have identified functions, and its infection cycle is well characterized.
ΦX174 also has a useful feature for prompting a genome model. The end of the virus always has the same short sequence of bases. By giving the model a short prompt related to the start sequence, the researchers could push it toward producing something connected to ΦX174.
Filtering AI-generated genomes
Before testing any generated sequence in bacteria, the researchers narrowed the outputs with several conditions. The goal was to remove sequences that were obviously unsuitable while still allowing meaningful differences from the original virus.
Evo 1 and Evo 2 were given over 2 million additional bases from viruses that infect bacteria, then fine-tuned with sequences specific to Microviridae. The team also experimented with prompt length. Too much starting sequence caused the model to reproduce the rest of the genome; too little produced many unrelated sequences. Prompts of four to nine bases worked best.
The filtering rules included several checks:
- Outputs needed a spike protein at least 60 percent identical to the real one.
- Sequences shorter than < 4,000 bases or longer than > 6,000 were removed.
- Outputs with runs of the same base more than 10 bases long were discarded.
- Sequences with unusual GC and AT base-pair frequencies were excluded.
After this computer analysis, 302 proposed viral sequences remained. The researchers chemically synthesized 285 of them and inserted them into bacteria to test the result.
What worked, and what did not
Most of the tested sequences did nothing. But 16 inhibited the growth of E. coli, which suggested that they were functioning as viruses. Nine of those were the original AI outputs, while seven had acquired additional mutations after being inserted into bacteria.
The working viruses were both familiar and different. Overall, only 16 of the 285 tested outputs were viable, or 5.6 percent. But among the generated sequences most similar to ΦX174, with 98 percent sequence similarity and up, viability rose to 46 percent.
That pattern matters. It suggests that the model’s most successful designs stayed close to the known virus. Most viable viruses had the same set of genes as ΦX174, and none had any change in the DNA stretch where duplication of the virus’s genome starts.
The source article also notes why this is not surprising. Past studies found that even a single base change that alters one amino acid in one of the virus’s proteins had, on average, a 20 percent chance of inactivating the virus entirely. In other words, ΦX174 can be highly sensitive to genetic change.
Why the result matters
This work does not show an AI casually inventing a dangerous new virus from scratch. The generated viruses were closely related to an existing bacteria-infecting virus, and most proposed genomes failed.
But it does show that large genome models can produce complete viral genomes that sometimes work in cells. That is a meaningful shift from generating isolated protein-coding sequences or gene-like fragments.
The researchers argue that the field should start preparing for the possibility that a related AI system could one day design a virus that targets vertebrates. The source does not claim that Evo 1 or Evo 2 did this. The warning is about direction of travel: models trained on biological sequence are improving, and their outputs may eventually require stronger planning around safety and oversight.
For now, the clearest lesson is technical. DNA-based AI can capture enough biological structure to generate viable versions of a small bacteriophage, but the useful outputs remained tightly constrained by the biology of the original virus.