AI Designs Killer Viruses in Lab for First Time

What could possibly go wrong? Scientists have used Artificial Intelligence to design new viruses capable of killing cells right here in the lab. This moment marks the first time such technology successfully generated whole genomes, meaning the full set of genetic instructions needed to build a working organism. Supporters claim this work offers hope for developing new treatments. Critics warn it raises urgent safety and security concerns immediately.

Researchers at Stanford University in California used the tool to create a genome for a virus that infects bacteria. The AI suggested thousands of different genomes. The team then created 302 of them in the lab before exposing them to bacteria. Overall, 16 of the viruses suggested by the AI were able to kill E.coli. These creations are bacteriophages, which only infect bacteria and cannot infect human, animal or plant cells.

Dr Brian Hie, a chemical engineer behind the research, revealed the results with these words: 'In this case, we wanted the model to generate the entire genome end-to-end in a single left-to-right pass.' The speed of discovery is terrifyingly fast. One wrong step could unleash something we cannot control. We must ask ourselves if our defenses are ready for what comes next.

We did not add anything." That was the firm stance from scientists who have now used artificial intelligence to design a brand new virus capable of infecting other cells. The study hit the pages of Science, published right alongside an urgent article warning about the dangers this advance could bring. Experts Dr Thomas Inglesby and Dr Maurice Hanke from Johns Hopkins wrote that while the technology holds promise for life sciences, it immediately raises serious biosafety and biosecurity questions. Their message was stark: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not."

To get here, the team utilized two specific tools called Evo1 and Evo2. These models function much like the large language model chatbots everyone knows, such as ChatGPT or Grok, but with a critical difference, they were trained on genetic codes instead of written text. Researchers fed the system data from two million genomes belonging to bacteriophages before asking them to invent new potential genomes. The scientists then synthesized these AI-created sequences in a lab and placed them into petri dishes containing E.coli bacteria. This forced the bugs to start making copies of the viruses immediately. Teams monitored the dishes closely to see if the bacteriophages had begun attacking and killing their bacterial hosts.

Samuel King, a PhD student working in the lab, described seeing clear spots appear on the plates as "extremely exciting." In their paper, the researchers noted that this work provides a blueprint for designing diverse synthetic bacteriophages and other useful biological systems at the genome scale. They pointed out that bacteriophages possess some of the smallest genomes known to science, which makes them far easier to create using current methods. However, they warned this is just a stepping stone toward using AI for much more advanced research down the line.

Dr Patrick Cai from the University of Manchester in the UK offered his own take on what it all means. "While these are relatively small bacteriophage genomes, the significance extends far beyond phages," he said. He explained that this suggests genome language models are beginning to learn the design principles encoded by evolution, effectively opening the door to AI-assisted genome writing. Tom Ellis, a professor of synthetic genome engineering at Imperial College London, called the achievement impressive but noted it highlights the real challenges involved in creating larger, more complex genomes later on.

"This is literally the smallest and easiest genome to make," Ellis told The Guardian. He added that an AI trained on dangerous pathogens could theoretically be used to design harmful viruses, yet he believes controlling access to genetic data and restricting the synthesis of risky genomes would help mitigate that risk, governments are already working on these measures. Still, he cautioned against overblowing the threat. "The threat from full AI design and writing of a genome of a virus or bacteria is very overblown," he stated. He argued that simply taking existing pathogens and making gain-of-function changes to their genomes is so much easier and much more likely to be a real pathogenic threat.

Gain-of-function research is essentially the scientific practice of genetically altering a pathogen to study how it might evolve, enhancing traits like transmissibility or virulence to better understand future pandemic threats. But the term became a lightning rod during the Covid pandemic, fueling fierce debate over whether such experiments at the Wuhan Institute of Virology played a role in the virus's origins. Some of those specific experiments were funded by US taxpayer dollars, adding another layer of complexity to the conversation as scientists push forward with these powerful new tools.