Large genome models have spent most of their short existence designing proteins. That was the logical starting point: proteins are the workhorses of biology, and understanding how to engineer new ones means gaining direct leverage over cellular chemistry. But a research team at Stanford University has now pushed these models into different territory, using them to generate complete viral genomes from scratch. The viruses they produced are real, functional, and in some cases structurally unlike anything found in nature.
From Protein Design to Full Genomes
The models at the center of this work, called Evo 1 and Evo 2, belong to a class of systems trained on DNA sequences rather than human language. The underlying logic mirrors that of large language models: instead of predicting the next word in a sentence, these systems predict the next nucleotide in a genetic sequence. DNA uses only four letters (A, T, C, and G), which might seem to simplify the task, but genomes are far from uniform. Some positions are biologically critical; others are almost irrelevant. A model that cannot distinguish between the two will produce sequences that look plausible but do nothing.
Evo 1 and Evo 2 appear to have learned that distinction, at least within certain domains. Earlier work showed they could output DNA sequences encoding functional proteins in bacteria and mimic gene structures found in complex cells. The Stanford team’s contribution was to ask whether the same models could generate an entire viral genome, not just individual genes.
The Virus They Chose, and Why
The test subject was a bacteriophage called ΦX174, a virus that infects E. coli and cannot infect humans. It is a deliberate choice on multiple levels. ΦX174 is small and well understood: 11 genes spread across roughly 5,400 bases, with every gene’s function already identified and the full infection cycle well characterized. It also ends with a consistent short sequence of bases, which gave the researchers a natural prompt. Feed those terminal bases to the model, and it should recognize the context and generate something related.
Before running experiments, the team fine-tuned Evo 1 and Evo 2 with more than 2 million additional bases of bacteriophage DNA, then narrowed the training further to sequences from Microviridae, the broader family ΦX174 belongs to. They then tested different prompt lengths and found that four to nine bases of the starting sequence produced the most useful outputs.
Not every output was worth testing. The team applied a series of filters: any proposed genome missing or severely degrading the spike protein (the structure the virus uses to latch onto bacteria) was discarded, as were sequences that were too long, too short, or showed unusual patterns in their base composition. After filtering, 302 candidate sequences remained. Of those, 285 were chemically synthesized and inserted into bacteria.
The result: 16 of the 285 sequences inhibited bacterial growth, indicating they functioned as actual viruses. Nine were direct AI outputs; seven had acquired additional mutations after insertion. The overall viability rate was 5.6 percent. Among sequences with at least 98 percent similarity to the original ΦX174, viability climbed to 46 percent.
What “Functional but Distinct” Actually Means
Here is what most coverage of this work misses. The interesting result is not that the AI produced viruses similar to ΦX174. It is that some of the viable viruses were genuinely different in ways that would be difficult to reach through ordinary evolution.
ΦX174 is notoriously fragile. Past studies have shown that a single amino acid change in any of its proteins carries, on average, a 20 percent chance of inactivating the virus entirely. That fragility means natural evolution tends to keep the genome locked in place. The AI, operating outside evolutionary pressure, explored combinations that biology rarely reaches. One of the viable viruses lost an entire viral protein and compensated with changes elsewhere in the genome. Another added a completely new gene. One replaced a ΦX174 gene with a gene from a distantly related virus.
These are not minor variations. They represent structural solutions that natural selection, constrained by the need to keep each intermediate step viable, would struggle to find.
The researchers were careful about safety. Evo 1 and Evo 2 were deliberately not trained on sequences from viruses that infect complex cells, including vertebrates. The logic is straightforward: even outputs that cannot be understood might be dangerous, so the training data was restricted before the question could arise. But the Stanford team also noted something worth sitting with: the same general approach, applied to a model trained on vertebrate-infecting viruses, could in principle produce something far more concerning. They suggest the field should begin thinking about that scenario now, before it becomes urgent.
In Short
Large genome models can now generate complete viral genomes that are functional, structurally novel, and in some cases unlike anything produced by natural evolution. The work was conducted carefully, using a bacteriophage that poses no human risk, and with deliberate restrictions on training data. The broader significance is not the specific viruses produced. It is the demonstration that AI systems can navigate the combinatorial complexity of an entire genome and arrive at solutions that biology, left to its own devices, would rarely find. That capability will matter enormously for medicine and bioengineering. It also raises questions that the field has not yet fully answered.
Based on reporting from Ars Technica.