Researchers at Stanford University and the Arc Institute used artificial intelligence to write complete viral genomes that had never existed in nature, built them in a laboratory, and found that 16 produced functioning viruses capable of infecting and killing E. coli bacteria.
The work, published in Science on Aug. 6, is the first demonstration of generative AI designing complete functional viral genomes. It was led by Brian Hie, an assistant professor of chemical engineering at Stanford and an innovation investigator at the Arc Institute, with Stanford bioengineering graduate student Samuel King.
What was built matters as much as that it worked. These are bacteriophages, viruses that infect bacteria and cannot infect humans, animals, or plants.
The Systems and the Constraints Applied
Two genome language models, Evo 1 and Evo 2, generated the sequences. They differ from chatbot-style systems in what they were trained on: genetic sequence data rather than written text.
Sequences from viruses capable of infecting complex organisms were deliberately excluded from the training data, which the team describes as a risk-limiting measure taken at the design stage rather than afterward. The work was conducted in a secure laboratory using non-pathogenic hosts.
Rather than producing individual genes or proteins, the models output entire genome sequences. Given short prompts drawn from the start of the natural ΦX174 sequence, the models wrote the rest. They generated roughly 700,000 candidate genomes. Researchers selected 302 for synthesis, successfully built 285, and introduced them into bacteria. Sixteen produced viable phages, a hit rate of about 5.6%.
Nine of the 16 matched the sequences the models produced. Seven acquired additional mutations after insertion into bacteria. Several of the designed phages showed faster lysis or higher fitness than the natural reference, and a cocktail of them overcame resistance in E. coli strains that a comparable mixture of naturally sourced phages could not.
The scale involved is worth stating precisely. ΦX174 is 5,386 bases. The SARS-CoV-2 genome is roughly 30,000 bases, about six times longer. Tom Ellis, a professor of synthetic genome engineering at Imperial College London, noted that the experiment targeted what he called the smallest and easiest genome to make.
The Medical Case for Doing This
Bacteriophage therapy is not hypothetical. Clinics in several countries already use phages for bacterial infections that no longer respond to antibiotics.
The obstacle is matching. Phages are highly specific, and finding one suited to a particular resistant strain, at the moment a patient needs it, is slow and uncertain. Clinicians often search collections and libraries hoping something in stock happens to work.
A method for designing a phage around one specific pathogen would change that timeline. With more than 2.8 million drug-resistant infections occurring annually in the United States, and an antibiotic pipeline that has struggled to produce genuinely novel agents, an alternative route has obvious appeal.
None of this is available to patients. The study reported no animal or clinical testing. Designing a phage that kills E. coli in a laboratory dish is separated from treating a person by safety testing, manufacturing standards, regulatory pathways, and clinical trials that do not yet exist for this approach. MedicalDaily has covered preclinical work restoring antibiotic effectiveness and an approval that took years to arrive.
The Gap the Biosecurity Warning Identifies
A Perspective by Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security ran in the same issue of Science, describing urgent biosafety and biosecurity questions.
Their argument is specific rather than general alarm. The companies that synthesize DNA to order screen customers and sequences voluntarily. No US law requires either check. And the screening that does occur works by sequence similarity, comparing an order against known dangerous sequences.
That architecture is built for sequences resembling something that exists. An AI-generated genome matching nothing in nature can pass a similarity check precisely because it resembles nothing on file. The two fixes the commentary names are a legal duty for synthetic DNA providers to screen every order and customer, and detection tools tuned to catch genomes matching nothing in nature. Neither exists in required, deployed form.
Evo 2 is openly available free of charge, which sharpens the question. The study authors say future whole-genome design work should involve safety and security professionals from start to finish.
Not everyone shares the alarm. Ellis argued the biosecurity threat from full AI genome design is overblown, because modifying existing pathogens remains far easier and poses a more likely threat than designing one from scratch. Inglesby and Hanke acknowledged that this team engaged with biosafety and biosecurity questions more deliberately than most developers of powerful biological AI models, while framing the question as no longer whether generative viral genome design will exist but how governance responds.
The Realistic Reading for Readers
Nothing here poses a risk to anyone's health today. The viruses built cannot infect people, and the study did not create anything capable of causing human disease.
The concern is about capability and the pace at which oversight follows it. That is a governance question for regulators, funders and synthesis companies rather than something individuals act on.
For patients facing drug-resistant infections, the practical route remains what it was: infectious disease consultation, culture-guided antibiotic selection, and in some cases compassionate-use phage therapy through the small number of centers that offer it under FDA expanded access. Anyone interested should ask their infectious disease physician rather than pursuing it independently, since phage therapy outside a regulated pathway is not available and products marketed directly to consumers are not evaluated.
The next developments will come from regulators and DNA synthesis companies rather than from the laboratory bench. MedicalDaily has covered other laboratory findings years from the clinic.
Key Questions Answered
What did the researchers build? Complete viral genomes designed by AI. The models generated roughly 700,000 candidates; 302 were selected for synthesis, 285 were built, and 16 produced functioning bacteriophages that infected and killed E. coli.
Can these viruses infect humans? No. Bacteriophages infect bacteria only. Sequences from viruses capable of infecting complex organisms were deliberately excluded from the training data, and the work used non-pathogenic hosts.
How well did the method work? About 5.6% of built genomes produced viable phages. Nine of the 16 matched the model output exactly and seven acquired mutations. A cocktail of the designed phages overcame resistance that natural phages could not.
Why is this medically interesting? Phage therapy is already used for antibiotic-resistant infections, but finding a phage matched to a specific resistant strain is slow and uncertain. Designing one to order could shorten that timeline.
Is this available as a treatment? No. The study reported no animal or clinical testing, and the approach is separated from patients by safety testing, manufacturing standards, and regulatory pathways that do not yet exist.
What is the biosecurity concern? DNA synthesis companies screen orders voluntarily, with no US legal requirement, and screening works by comparing sequences to known dangerous ones. A sequence matching nothing in nature can pass that check.
Does everyone agree it is dangerous? No. One synthetic genome specialist argued the threat is overblown, since modifying existing pathogens is easier and more likely than designing one from scratch, and noted the target genome was the smallest and easiest available.
