AI Designs First Fully Functional Synthetic Viruses

Researchers at Stanford University and the Arc Institute used the genome language models Evo 1 and Evo 2 to design functional bacteriophage genomes from scratch, with 16 of roughly 300 synthesized candidates proving viable against antibiotic-resistant E. coli. The development signals a new phase for AI-driven biological design that carries implications for pharmaceutical innovation, biosecurity oversight, and regulatory frameworks worldwide.  

Scientists at Stanford University and the Arc Institute report that AI has, for the first time, designed complete viral genomes that function in living cells. In a study published in the journal Science on Aug. 6, 2026, researchers used the genome language models Evo 1 and Evo 2 to generate nearly 700,000 candidate bacteriophage genomes, of which 16 proved capable of infecting and destroying antibiotic-resistant Escherichia coli.  

The team deliberately narrowed the AI models' focus to bacteriophage Phi X-174, a virus that infects only E. coli and poses no threat to humans, along with roughly 15,000 related phages, while excluding any genetic data tied to organisms capable of infecting humans, animals, plants, or fungi. The precaution reflects the dual-use tension inherent in the technology. Moritz Hanke, Researcher, Johns Hopkins Center for Health Security, notes that a genome language model could in principle be asked to design an influenza genome modified to be more infectious or more lethal, underscoring why the researchers built safeguards into both the training data and the experimental scope from the outset.

Bacteriophages, viruses that attack bacteria rather than human cells, have drawn renewed interest as a potential complement to antibiotics amid rising antimicrobial resistance, a threat the World Health Organization has linked to nearly 5 million deaths globally in a single recent year. Historically, phage therapy has been limited by the difficulty of finding and engineering strains effective against specific, evolving bacterial targets. Evo 2, trained on billions of DNA sequences spanning multiple domains of life, was designed to learn the patterns underlying functional genomes rather than simply predicting the next letter in a sequence, giving researchers a tool to generate novel viral candidates instead of relying solely on those found in nature.

AI is already reshaping how drug candidates are identified and designed. For example, AI-assisted biomarker discovery is being applied to precision oncology, helping researchers identify molecular targets and design therapies earlier in the development process. 

Of the approximately 300 AI-generated genomes the researchers synthesized and tested in secure laboratory conditions, 16 produced viable phages capable of replicating inside bacterial cells, with some outperforming the naturally occurring Phi X-174 against resistant bacterial strains. Brian Hie, Assistant Professor, Stanford University, says that the team wanted to be extremely careful in scoping the project, while co-author Samuel King describes the move into virus design as a logical next step for genome language models.The researchers have made Evo 2 openly available, a decision intended to support legitimate research into pathogens and antibiotic alternatives.

In Mexico, a Senate commission has spent recent months drafting a horizontal, risk-based framework intended to integrate AI principles across sectoral laws, including health regulation, though no general AI statute has yet entered into force. As COFEPRIS separately works to modernize clinical research and biotechnology approval pathways, the emergence of AI tools capable of generating functional biological agents adds a new dimension to that regulatory agenda, one that stakeholders in Mexico's pharmaceutical, biotechnology, and life sciences sectors are likely to watch closely in the coming months.