Generative AI has reached a milestone that researchers have anticipated for years. Scientists at Stanford University and the Arc Institute have successfully used AI to design complete viral genomes from scratch, with 16 of those designs confirmed as fully functional, biologically active viruses capable of infecting bacteria.

According to a preprint released on bioRxiv on September 12, 2025, this represents the first time generative AI has produced working viral genomes end-to-end. The viruses in question are bacteriophages — viruses that target bacteria, not humans.

How the AI-Designed Phages Were Built

The research team, led by Brian Hie at Stanford with collaborators from the Arc Institute, NVIDIA, and UC Berkeley, employed genome language models known as Evo 1 and Evo 2. Rather than predicting the next word in a text sequence, these models predict the next base pair in a DNA sequence. This approach mirrors the transformer architecture behind large language models like those powering ChatGPT, but applied to the four-character alphabet of DNA — A, C, G, and T.

Using the architecture of ΦX174 — a naturally occurring bacteriophage that infects E. coli — the team generated 302 candidate phage genomes. Of those candidates, 16 were experimentally verified to assemble into viable viruses capable of infecting and lysing E. coli bacteria.

Several AI-designed variants demonstrated replication advantages of up to 65 times compared to the natural template. The researchers also evaluated these phages in cocktail therapies against antibiotic-resistant bacterial strains, where they exhibited significant efficacy.

Antibiotic Resistance as a Driving Force

Antibiotic resistance remains one of the most pressing yet underrecognized crises in global healthcare. The World Health Organization has identified antimicrobial resistance as one of the top ten global public health threats, with drug-resistant infections contributing to millions of deaths annually. Phage therapy, which deploys viruses to eliminate specific bacteria, has existed conceptually since the early 20th century and was widely used in parts of the Soviet Union before the antibiotic era. However, it has never achieved widespread clinical adoption in Western medicine, largely because engineering the appropriate phage for a given bacterial strain is a laborious and time-intensive process. A handful of phage therapy cases have been treated under FDA compassionate-use provisions, but scalable, on-demand phage design has remained out of reach.

The Evo models were trained on vast genetic datasets encompassing trillions of DNA base pairs, building on prior work introduced in 2024. The natural ΦX174 phage has a genome of approximately 5,386 base pairs — remarkably small by biological standards.

Biosecurity Considerations

The 16 functional phages were designed within the parameters of a known natural template and specifically target E. coli. The researchers emphasized that they were not creating novel pandemic pathogens. Nevertheless, the underlying capability — a model that can construct working viral machinery — raises governance questions that warrant careful discussion as the technology progresses. The dual-use nature of genome-scale generative models places this work within an ongoing policy conversation about oversight frameworks for AI systems capable of designing biological entities, a topic that has drawn attention from both U.S. biosecurity agencies and international governance bodies.

NVIDIA's participation in the collaboration is also notable for those monitoring AI infrastructure investment trends. Genomic modeling at this scale demands substantial computational resources, and the life sciences sector has emerged as a significant application domain for high-performance GPU clusters. The work follows a broader trajectory in which AI-driven biology — from DeepMind's AlphaFold for protein structure prediction to whole-genome language models — has become one of the most computationally intensive and commercially consequential frontiers in applied AI.