Researchers have used generative AI to design synthetic viruses that don't exist in nature - and some of them successfully attacked antibiotic-resistant bacteria in lab tests. The work, published in Science, points toward a future where AI speeds up the search for phages that could treat infections that no longer respond to antibiotics.
Bacteriophages are viruses that infect bacteria, not people. The research team used two AI models, Evo 1 and Evo 2, to generate new phage designs similar to Phi X-174, a well-studied virus that attacks E. coli. According to Smithsonian Magazine, the models produced roughly 700,000 candidate designs. The team selected 285 to build in the lab.
Sixteen of those synthetic phages stopped E. coli growth in Petri dishes. In follow-up experiments, the same 16 phages successfully attacked two antibiotic-resistant E. coli strains. Phi X-174 and a mix of natural phages failed those same tests.
The stakes are significant. The World Health Organization says bacterial resistance was linked to more than 4.7 million deaths globally in 2021.
AI does the heavy lifting
For years, researchers have explored phages as an alternative to antibiotics for drug-resistant infections. The challenge has been finding phages that work reliably. The allure of this approach is speed: AI can generate designs in hours that would take months to isolate naturally.
Isaac Bogoch, an infectious diseases specialist at the University of Toronto who was not involved in the research, told Al Jazeera: "AI-designed viruses could have some potential benefits, such as the creation of targeted bacteriophages that could possibly help us tackle antibiotic-resistant infections in new ways."
The technology isn't there yet for broad application. Tom Ellis, a synthetic genome engineer at Imperial College London, pointed out that Phi X-174 has an extremely small genome. More complex viruses are far harder to design with current models.
The field of Generative AI and LLM technology is advancing quickly, but its application to biology is still early-stage.
Safety measures built in
The researchers said they left certain genomes out of Evo's training data to prevent the system from generating viruses that could infect humans, animals, plants, or fungi. They also applied safety measures beyond standard research practice and documented those steps as a proposed biosafety framework.
Because any technology that can create useful viruses could be misused on dangerous ones, the team published its containment protocols for other labs to follow.
For practitioners in AI for Science & Research, the payoff could be significant. One direction the team is pursuing is mixed phage therapy - pairing genetically distinct phages so that bacteria would face a steeper challenge to evolve resistance around them.
Why this matters for AI for Science & Research
This study demonstrates that generative models can produce working biological designs, but there's a gap between initial hits and clinical applications. The 16 phages that cleared preliminary tests still need to move from the discovery phase to validation, animal testing, and eventually human trials.
For scientists and researchers, the takeaway is this: AI-driven discovery is now producing results that can be physically tested, but the bottleneck has shifted from generating candidates to screening them. The field needs automated ways to evaluate thousands of AI outputs if it hopes to scale the approach from E. coli to more complex pathogens.
Your membership also unlocks:
