A study from Stanford University and the Arc Institute, published in Science, reported that AI designed viral genomes not found in nature and turned some of them into viable, self-replicating viruses. In the experiment, the Evo model generated 700,000 candidate viral genomes. Researchers selected 285 for synthetic DNA construction, and 16 ultimately produced new viruses capable of infecting bacteria.
Evo generated 700,000 candidate genomes
The paper is listed at https://www.science.org/doi/10.1126/science.aec2657. The model used in the work is called Evo and was built by the Arc Institute. The report describes it as a generative model for biological sequences trained on about 9 trillion nucleotides spanning animals, plants, microbes, and viruses.
The underlying material was DNA sequence data. As described in the source article, DNA uses four nucleotide “letters” — A, C, G, and T — and genes are built from those letters in ordered combinations. Evo was presented as learning those patterns directly from massive sequence data rather than from hand-written biological rules, then producing 700,000 candidate viral genomes in one run.
285 were synthesized and 16 survived
Researchers narrowed the 700,000 candidates to 285 sequences, converted them into synthetic DNA molecules, and introduced them into Escherichia coli for testing. Most culture plates showed no obvious effect and the bacteria continued to grow. On some plates, however, distinct plaques appeared, indicating that viral infection and replication had taken place.
The final result was 16 new viruses that, according to the article, were alive in the experimental sense: they could infect bacteria and self-replicate. The report describes this as the first time AI has designed a complete living genome from scratch.
Some AI-designed viruses beat natural ΦX174 in lysis speed
The study did more than show that the viruses could function. It also examined how they performed. According to the article, some of the AI-generated viruses reproduced and lysed bacteria faster than the natural reference virus, ΦX174.
Cryo-electron microscopy also captured a structural detail. One virus, named Evo-Φ36, used a DNA-packaging protein on its shell that came from a more distantly related evolutionary source. The article framed that as evidence that the AI design process combined biological parts from different origins while still producing a working virus.
Why ΦX174 matters
The virus redesigned in the study was ΦX174. The article says it was recovered from Paris sewage in 1935, infects only E. coli, and is harmless to humans and animals. That made it a long-standing model organism in molecular biology.
Its historical role is even larger. In 1977, the Sanger team completed the first full genome sequencing in history using ΦX174. The article draws a direct line from that moment — the first full reading of an organism’s instructions — to the new work 49 years later, which it describes as the first full writing of one by AI, again in the same species.
The source also notes that ΦX174 has a genome of about 5,000 bases and 11 genes. Despite its small size, those genes are nested and overlapping, which makes the virus a difficult design target.
Brian Hie on generative genome design
Brian Hie, a Stanford assistant professor who led the research, was quoted in the article as saying: “Welcome to the era of generative genome design.” The line reflects the shift described in the report: AI is not only reading and analyzing biological sequences, but also taking part in the design and experimental validation of new genomes.
Resistance test: AI phage cocktail broke through three strains
The article connects the work to antibiotic resistance. It cites a Lancet GRAM project estimate that antibiotic resistance will directly cause 39.1 million deaths between 2025 and 2050. One alternative approach in medicine is phage therapy, which uses viruses to kill bacteria, but bacteria can develop resistance to phages as well.
In this experiment, the team first cultivated E. coli that had become fully immune to natural ΦX174, then tested two cocktail approaches:
- a natural phage mixture, which did not win;
- an AI-generated phage mixture, which quickly broke through three resistant strains.
The article presents that result as the first proof that AI can keep up with bacterial evolution by generating new phages at scale.
From reading life to writing it
The source places the study on a longer scientific timeline: in 1977, humans fully read a genome for the first time; now, AI has written a new viral genome that can survive and self-replicate. The material cited in the original article came from the WeChat public account Xinzhiyuan, credited to ASI Qishilu.