Researchers have used artificial intelligence to design hundreds of new versions of a virus from scratch, with 16 of 285 AI-generated genomes producing working phages capable of infecting E. coli. The experiment marks a shift in what AI can do in biology - it has moved beyond analyzing genetic code to writing entirely new versions that can function in the real world.
The work, led by researchers using two genome-focused AI models called Evo 1 and Evo 2, demonstrates that machine learning can generate complete viral genomes that actually work. Most of the designs failed, but the ones that succeeded behaved in ways comparable to the original virus despite having substantially different DNA sequences.
What the AI did
The viruses in question belong to a relatively simple class called phages, which infect bacteria rather than humans or animals. They are plentiful and highly adapted to their hosts. When a phage finds the right bacterium, it attaches and injects its genetic material, taking over the cell's machinery to produce new copies until the host bursts.
The researchers trained the AI on more than 2 million phage genomes, then fine-tuned it using about 15,000 close relatives of their target, a well-studied phage called ΦX174. With 5,400 DNA letters and 11 proteins, ΦX174 serves as a streamlined test subject - its genome is small enough to synthesize quickly and its outcome is clear: either the DNA produces a functioning phage or it doesn't.
The team chemically synthesized the AI-designed sequences and introduced them into E. coli. Sixteen produced working phages. One striking outcome: a combination of genetic elements that wouldn't have worked in the original phage turned out to function efficiently in the redesigned version.
Resistance testing and possible treatments
There's a practical angle to this work beyond the lab demonstration. The researchers exposed their AI-generated phages to an E. coli strain that could resist the original virus. After repeated rounds, hybrid phages emerged that could infect the previously resistant bacteria. The AI didn't directly design those final phages - it generated a diverse starting population, giving evolution more possibilities to work with.
This approach could play a role in the fight against antibiotic resistance, where bacteriophages have been investigated as alternatives to drugs because they target only specific bacteria while leaving human cells alone. The main challenge is finding the right phage for the right bacterial infection, and AI could help researchers predict which strains will work against stubborn drug resistant bacteria.
The field of AI for Science & Research has been building toward this kind of capability for some time.
The study is, however, a long distance from designing phages for patients. This particular microbe was unusually simple, with just 11 proteins, and many phages that could fight serious infections have far larger genomes. Any medical application would require the basic tests, production standards, and lengthy approval work for safety and efficacy - not to mention difficult regulatory questions around personalized phage designs for individual patients. The team behind this work treats it as a demonstration of what might be possible, not the arrival of AI-designed phage therapy.
Researchers have already spent decades studying phages to understand basic genetics. Now those same small structures are helping scientists test whether AI can design a complete genome that actually works, rather than merely perform predictive analyses on existing DNA - central to the evolving toolkit in AI for Microbiologists, where new design tools are joining standard lab methods.
Waiting questions and biosecurity
The open question is whether this approach can scale beyond a simple blueprint. Can AI design more complex phages, handle bacteria that cause serious human infections or drive reliable results outside the lab? No. The most important part - and what the researchers emphasize - is that none of this yet indicates AI can simply create dangerous viruses on demand.
It does, however, confirm that AI systems are beginning to generate working biological sequences instead of just analyzing them. The findings make it increasingly important to consider how such systems are being trained, what biological information is available to them, and how risky designs should be screened for biosecurity.
These questions aren't for research laboratories alone. Biologists, regulators, ethicists and biosecurity experts will all need to weigh in as the technology develops.
Why this matters for science and research professionals
For scientists working at universities, biotech companies or pharmaceutical firms, the bottom line is concrete: generative genome design is now at the edge of laboratory validation, and a team has shown it can work. The technology is simple - and it is not yet ready for clinical use or biosecurity applications - which means the immediate challenge is validating these methods on more complex biological systems.
Professionals who track these tools should watch whether these models extend their work to larger pathogen genomes: the scientific year to pay attention to this line of research is just starting, and the underlying struggle - whether AI can design code that survives outside the computer - has gained its first clear benchmarks.
Your membership also unlocks:
