Stanford's Evo 2 Writes Viral Genomes No Evolution Ever Produced — and They Work

For the first time, artificial intelligence has written the complete genetic code for a living, functioning virus — one that nature never evolved on its own. Researchers at Stanford University and the Arc Institute used a genome language model called Evo 2 to design 16 novel bacteriophages from scratch, all of which proved capable of infecting and killing Escherichia coli bacteria, including strains resistant to both standard antibiotics and natural viruses.

The work, led by Brian Hie and Samuel King at the Arc Institute and published in the journal Science after an initial preprint in September 2025, represents a watershed moment in generative biology. It demonstrates that a transformer-based model trained on DNA sequences can do far more than predict protein structures or flag mutations — it can write entirely new genomes that produce organisms which actually work.

How Evo 2 Writes Life

Evo 2 belongs to a new class of models called genome language models. Where GPT and Claude are trained on human text, Evo 2 was trained on DNA — specifically, on 9.3 trillion nucleotides drawn from more than 128,000 whole genomes spanning every domain of life: bacteria, archaea, plants, animals, and everything in between. Released in March 2025 as an open-source collaboration between the Arc Institute, NVIDIA, and academic partners, the model reads and generates DNA sequences at single-nucleotide resolution across context windows of up to one million base pairs.

The architecture is conceptually similar to an autoregressive language model: given a prefix of a DNA sequence, Evo 2 predicts the most likely next nucleotide (A, T, G, or C), base pair by base pair. But because DNA encodes the machinery of life, the sequences it generates are functional programs, not just strings. A well-written viral genome is a complete blueprint that, when synthesized and introduced into a host cell, can hijack the cell’s machinery to produce new virus particles.

Hie and King fine-tuned both Evo 1 and Evo 2 on existing bacteriophage genomes — particularly ΦX174, a well-studied virus that infects E. coli. They then asked the models to generate novel genome-length sequences of roughly 5,000 to 7,000 nucleotides, each a candidate blueprint for a brand-new phage.

From 300 Designs to 16 Living Viruses

The researchers synthesized and tested nearly 300 candidate phages generated by the models. Of these, 16 produced viable viruses capable of replicating inside E. coli C, the non-pathogenic laboratory strain used in the experiments. That is a roughly 5% success rate for generating entirely new, functional viral genomes from scratch — a rate that, while modest, would have been unthinkable just two years ago.

The significance goes beyond raw numbers. Several of the AI-designed phages outperformed the natural ΦX174 reference virus in their ability to infect and kill bacteria. Even more striking, when the researchers pitted the synthetic phages against E. coli strains that had evolved resistance to natural bacteriophages, the AI-designed variants were able to overcome that resistance — killing bacteria that no known natural phage could touch.

A cocktail combining multiple AI-generated phages proved even more potent, suggesting a path toward phage therapy regimens that can outmaneuver the bacterial defense mechanisms that routinely defeat single-phage treatments.

Why This Matters: The Antibiotic Resistance Crisis

The World Health Organization classifies antimicrobial resistance as one of the top global public health threats. Drug-resistant bacteria currently kill an estimated 1.27 million people annually, and the toll is projected to rise sharply as the pipeline of new antibiotics dwindles. Phage therapy — using viruses that exclusively target bacteria — has been pursued for over a century, particularly in Eastern Europe, but has been limited by the difficulty of finding or engineering phages that work against specific resistant strains.

Evo 2’s results suggest a fundamentally new approach: rather than hunting for the right phage in nature, researchers can now design them. Because the model has internalized the statistical grammar of viral genomes across the entire tree of life, it can propose sequences that exploit bacterial vulnerabilities no naturally occurring phage has happened upon. The 16 functional phages from this study are proof that AI-designed viruses can clear a barrier that natural phage discovery often cannot.

The Biosecurity Elephant in the Room

The same capability that could fight superbugs could, in principle, be used to design pathogens. The researchers and outside commentators have been unusually direct about this tension. A Scientific American analysis noted that the study “confirmed a biosecurity gap” — the gap between the accessibility of powerful generative biology tools and the oversight frameworks needed to govern them.

Evo 2 is open-source and openly downloadable. Anyone with sufficient technical knowledge can run it. The Arc Institute has argued that openness accelerates beneficial research and enables defensive applications, and the phage work itself is a prime example. But the dual-use implications are clear: if an AI model can write a genome for a virus that kills bacteria, the same methods could potentially be directed toward designing agents that harm humans, animals, or crops.

The study’s authors have called for robust safeguards, including DNA synthesis screening — automated checks that gene synthesis companies perform before manufacturing ordered sequences — and responsible-use norms for generative biology models. Whether the current patchwork of voluntary screening will prove adequate as these models grow more capable remains an open and urgent question.

Beyond Phages: A General Engine for Biological Design

The phage results are a milestone, but they may be the tip of the iceberg. Evo 2’s generative capabilities extend across all domains of life. The Arc Institute has already demonstrated that the model can identify disease-causing mutations in human DNA, predict the effects of genetic edits, and generate plausible sequences for CRISPR-associated genes and other functional elements.

In a March 2026 retrospective titled “Evo 2: One Year Later,” the Arc Institute team described the phage work as the first concrete proof that genome language models can move from prediction to full-scale design — from reading the code of life to writing it. Subsequent work is exploring applications in agricultural biotechnology, where designer phages could protect crops from bacterial blights, and in industrial biomanufacturing, where engineered microbes need tailored viral defenses.

The broader trajectory is clear. Language models began by generating text. They progressed to code, then to images, audio, and video. Now they are generating the code of life itself. The Stanford and Arc Institute results show that this capability has crossed the threshold from speculation to working reality — and that the scientific, medical, and governance communities will need to move quickly to keep pace.

Sources

The findings reported here draw on peer-reviewed research published in Science, supplementary materials from the Arc Institute and Stanford University, and reporting from Nature, Scientific American, Genetic Engineering & Biotechnology News, and the Australian Broadcasting Corporation. The underlying preprint is available on bioRxiv.