A Stanford team has accomplished what recently seemed like science fiction: generative AI models have, for the first time, designed complete viral genomes that proved viable and successfully replicated under laboratory conditions. The results of this work, published in the prestigious journal Science, open a new era in synthetic biology.
How it works
At the core of the experiment are the Evo1 and Evo2 models, trained on vast arrays of genetic data—from viruses and bacteria to plants and humans. The operating principle resembles large language models, but instead of predicting words, they predict DNA sequences. After refinement, the algorithms were aimed at generating bacteriophages—viruses that attack only specific types of bacteria.
From thousands of generated variants, the researchers selected 302 of the most promising ones and synthesized them in the laboratory. The result is impressive: 16 of them demonstrated effectiveness against E. coli, with some even overcoming the bacteria's natural defense mechanisms. Notably, the created sequences have no analogs in nature—these are fundamentally new biological constructs.
A moment of triumph and safety questions
Graduate student Samuel King recalls how the team discovered the first signs of phage activity early in the morning: transparent spots appeared on the Petri dishes—a clear marker of viral replication. When the results were shown to the entire group, the laboratory literally erupted in applause, confirms Assistant Professor Brian Hie.
However, the scientific community is divided in its assessment. On one hand, the development of new bacteriophages opens prospects for combating infections resistant to antibiotics. On the other, experts from the Johns Hopkins Center for Health Security, Thomas Inglesby and Moritz Hänke, are raising the alarm, speaking of "urgent biosecurity questions." Hänke warns: "You could say, 'Hey, genomic language model, make me a flu genome modified to be more contagious or more lethal.'"
The researchers took precautions: viruses that infect complex organisms were excluded from the training database, and the work was conducted in a secure laboratory. Nevertheless, the discussion about how far one can go in designing biology is just beginning.
A look to the future
Skeptics, like Tom Ellis from Imperial College London, remind us that the AI-created viruses are "literally the smallest and simplest genomes" (about 5,400 base pairs), whereas the minimal genome of a living cell is roughly 500,000, and the human genome is three billion. Creating a dangerous pathogen would require far more complex work.
Marc Güell from Pompeu Fabra University called this "a very significant turning point," since "for the first time in history, people are beginning to design biology on a computer." And this is indeed true: we stand on the threshold of an era where AI becomes not an analysis tool, but a full-fledged creator of biological systems.
My analysis: this breakthrough is a double-edged sword. On one hand, synthetic phages could become a salvation in the age of antibiotic resistance. On the other, we are only beginning to understand the risks associated with delegating control over the fundamental processes of life to AI. The biotech market and the crypto industry are united here by one thing: the speed of innovation adoption always outpaces regulatory frameworks, and we need proactive safety mechanisms, not reactive ones.