AI Created 16 New Viruses: It Learned from 9 Trillion Nucleotides
The Evo model designed genomes never seen in nature, and 16 of them grew into functional viruses. Scientists warn that regulation is lagging behind the technology's capabilities.
The Evo model designed genomes never seen in nature, and 16 of them grew into functional viruses. Scientists warn that regulation is lagging behind the technology's capabilities.
Researchers at Stanford and the Arc Institute have made a frightening yet potentially revolutionary breakthrough: artificial intelligence has successfully designed 16 entirely new viruses that do not exist in nature. The model, named Evo, was trained on a massive amount of genetic data, learned DNA patterns, and then used them to create functional organisms, according to Tom's Hardware, citing a report from The New York Times.
The study, published in the journal Science, shows that out of 285 genomes proposed by the AI and tested, 16 were successfully assembled into viruses capable of infecting bacteria and replicating. Although the newly created viruses are harmless to humans and exclusively attack the E. coli bacterium, the experiment raises critical questions about the safety and necessity of regulating artificial intelligence in biology.
Evo was trained on more than 9 trillion nucleotides, the building blocks of DNA, encompassing 128,000 genetic sequences from millions of animals, plants, microbes, and viruses. Like a language model that learns which word combinations make sense, Evo discovered statistical patterns in the arrangement of nucleotides that result in biologically meaningful sequences. In other words, it deciphered the 'grammar' by which nature assembles DNA.
The scientists then focused the model on the simple virus Phi X-174, which has been studied for over 100 years and contains only 11 genes and 5,386 nucleotides. After analyzing this virus and about 15,000 of its closest relatives, Evo proposed as many as 700,000 new potential genomes. The research team narrowed the selection to the 285 most promising ones, chemically synthesized them, and observed the results.
To demonstrate potential applications, the scientists first developed three strains of E. coli resistant to the natural Phi X-174 virus. When these resistant bacteria were exposed to cocktails of viruses designed by the AI, the phage populations evolved during the experiment and ultimately overcame resistance in all three strains. This result hints at applications in the fight against antibiotic-resistant superbugs.
The study's authors emphasize that they were extremely cautious: they did not train the model on data from organisms that can infect humans, and the new viruses are very similar in composition to the original, only with a different arrangement of mutations. However, the very fact that AI can autonomously design living organisms raises concerns.
Experts warn that such applications far outpace existing protective mechanisms and regulations. Concerns are further fueled by earlier incidents with artificial intelligence, such as when an OpenAI agent hacked the Hugging Face platform on its own initiative. Although the current study is a proof of concept on harmless viruses, scaling the results to pathogens that attack humans would require a completely different level of oversight, and that is precisely what is currently lacking.
"AI creates 16 new viruses that never existed in nature after learning DNA patterns from 9 trillion nucleotides, experts warn that such applications are far ahead of the necessary protective measures," reads the headline of Tom's Hardware's report.
For comparison, the human genome contains more than 3 billion building blocks, making this experiment with a simple virus of just a few thousand nucleotides only the beginning. The scientific community now faces the challenge of harnessing the immense potential of this technology while preventing its misuse before it's too late.
