Stanford researchers used the Evo 2 generative-AI model to design and synthesize almost 300 novel bacteriophages. Laboratory tests confirmed that 16 of those viruses efficiently kill E. coli. The work shows that de novo biological design can jump from a line of code to a living antimicrobial agent.
From a Digital Sequence to a Real Virus
Assistant Professor Brian Hie, together with the Dieter Schwarz Foundation Stanford Data Science Institute, fed Evo 2 the genome of the well-studied phage ΦX174. Evo 2 treats DNA like a language, predicting sequences that should fold into functional proteins and assemble into a viable virus. The model churned out hundreds of distinct DNA strings; chemical synthesis turned them into a library of nearly 300 phage particles.
Turning a computer-generated string into a physical virus required more than a printer-like step. The team ordered custom DNA oligonucleotides, assembled full genomes, and introduced each construct into a bacterial host that could “boot up” a phage. That workflow—design, synthesis, assembly, and rescue—has long been standard for natural phage isolates but has rarely been applied to wholly AI-crafted genomes.
Screening for Antibacterial Activity
The real test was whether any AI-designed phages could infect and destroy a bacterial target. The researchers challenged the 300 candidates with a laboratory strain of E. coli, a common cause of urinary-tract infections and a model for antibiotic-resistance studies. After plaque assays and growth-inhibition measurements, 16 phages stood out for rapid lysis of bacterial cultures and high burst sizes (the number of new phage particles released per infected cell).
Those 16 hits didn’t just survive the screen; they outperformed many natural ΦX174 relatives in killing efficiency. The data suggest Evo 2 can reproduce known viral functions and also explore sequence space that biology has never sampled, arriving at solutions that are both viable and potent.
Why It Matters for Antimicrobial Research
Phage therapy—using viruses that prey on bacteria—has resurfaced as a possible answer to the global rise in antibiotic-resistant infections. Traditional discovery relies on hunting for phages in sewage, soil, or animal microbiomes, a process that can take months and yields unpredictable results. Evo 2 offers a compute-first alternative: scientists specify a bacterial target, let the model generate candidate genomes, and then test a focused set in the lab.
If the pipeline scales, it could compress years of environmental sampling into weeks of in-silico design and high-throughput synthesis. Tailoring phage genomes on demand also opens the door to engineering extra traits, such as broader host ranges or resistance to bacterial anti-phage defenses.
Risks and Open Questions
The breakthrough is still early. All experiments have been confined to petri dishes; no animal models or clinical trials have been run. Safety concerns linger—introducing synthetic viruses into patients or the environment could have unforeseen ecological impacts. Regulatory frameworks for genetically engineered phages are still evolving, and the path from laboratory efficacy to approved therapy could be long.
Moreover, while Evo 2 succeeded with a relatively simple, well-characterized phage like ΦX174, it remains to be seen whether the model can handle larger, more complex viral genomes or design phages that must navigate bacterial immune systems in vivo. The current success may reflect the model’s training on abundant ΦX174 data rather than a universal capacity to generate any virus from scratch.
What Comes Next
The Stanford team will test the 16 lead phages against clinically relevant E. coli strains, including those resistant to multiple antibiotics. Parallel work will explore whether Evo 2 can be prompted to design phages for other problematic bacteria, such as Klebsiella or Pseudomonas. Scaling up synthesis—from a few hundred candidates to thousands—will also be a priority, as larger libraries could reveal even more potent or broadly acting viruses.
Beyond phage therapy, the methodology could be repurposed for other bio-engineering goals: designing enzymes for industrial chemistry, creating synthetic metabolic pathways, or even constructing novel vaccine vectors. Evo 2’s ability to generate functional DNA sequences points toward a future where “code” becomes a universal design language for living systems.
Takeaway
Stanford’s Evo 2 model has turned a computational prediction into a living antimicrobial tool, showing that AI-driven de novo biology can produce functional viruses capable of killing a major pathogen. The proof-of-concept narrows the gap between digital design and real-world therapeutics, but safety, regulatory, and scalability challenges must be solved before synthetic phages become a routine part of the antimicrobial arsenal.