In a demonstration of how far artificial intelligence has penetrated biology, researchers used a genomic AI to design 16 new viruses that never existed in nature. The viruses, built from AI-created DNA sequences, successfully infected bacteria and reproduced. The study published in Science reveals both the power and the potential peril of learning to write the language of life.

What You Need to Know

A team at Stanford University and the Arc Institute trained an AI model called Evo on over 9 trillion nucleotides from 128,000 genomes. After learning the statistical patterns of DNA arrangement, Evo generated new viral genomes. Of 285 AI-designed viruses tested, 16 assembled into functioning viruses that attacked E. coli bacteria. The researchers deliberately avoided training on human pathogens, but experts warn the technique could be adapted for dangerous applications without proper oversight.

How Evo Learned the Language of DNA

Evo is a genomic AI model trained on trillions of nucleotides, the building blocks of DNA found in every living organism. The model studied over 128,000 genetic sequences drawn from millions of animals, plants, microbes and viruses. It learned statistical patterns in how biological DNA is arranged, much like a language model learns which word combinations make sense.

  • Training scope: Evo analyzed 9 trillion nucleotides from over 128,000 genetic sequences.
  • Source diversity: Data came from millions of animals, plants, microbes and viruses.
  • Pattern discovery: The model learned which nucleotide combinations produce biologically viable sequences.

After mastering these patterns, Evo could design new genes that instruct cells to make specific proteins. The team wanted to test whether the AI could create a complete blueprint for a simple organism — a virus with only a few thousand building blocks. For context, the human genome contains more than three billion.

Designing New Viruses From Scratch

The researchers focused on the Phi X-174 virus, a well-studied bacteriophage that attacks only E. coli bacteria. They trained Evo on the 11 genes and 5,386 nucleotides of Phi X-174 along with about 15,000 of its closest relatives. The model then proposed 700,000 potential new genomes. The team trimmed that list to 285 candidates, synthesized them chemically and tested whether they could come to life.

Sixteen of those AI-generated viruses successfully assembled into functioning phages capable of infecting E. coli and reproducing. The new viruses closely resemble Phi X-174 but carry mutations in how their nucleotides are arranged. They do not pose a threat to humans because the team never included human pathogen data in the training set.

In a separate experiment, the researchers generated three E. coli strains resistant to natural Phi X-174. They then exposed those resistant bacteria to a cocktail of AI-designed phages. The phage populations evolved during the experiment and ultimately overcame resistance in all three strains, demonstrating a practical application for combating antibiotic resistance.

Implications for Biosecurity and Regulation

The study highlights both promise and peril. On the positive side, AI-designed viruses could help develop targeted therapies for bacterial infections. Researchers have long synthesized viruses to study them and test antiviral drugs. However, the Evo approach goes further by creating entirely new organisms without a natural template.

Experts worry that such capabilities are advancing faster than the guardrails meant to control them. The technology could theoretically be adapted to design pathogens that affect humans. The researchers say they were careful, but the possibility remains. They note that replicating the same results in human-infecting viruses would be far more complex. Still, the demonstration shows that the underlying method works.

Why This Matters

This experiment shifts the biosecurity landscape. For decades, creating new viruses required editing existing genomes. Now an AI can generate viable sequences from scratch after training on natural DNA patterns. The technique opens possibilities for synthetic biology, from drug development to environmental remediation. But it also lowers the barrier for creating novel pathogens. Policymakers and regulators face an urgent need to establish oversight for genomic AI models before the technology becomes widely accessible. The line between beneficial research and dangerous misuse has never been thinner, and the tools to cross it are already here.