Researchers have successfully created the first artificial intelligence-designed viruses, marking a significant advancement that fuels optimism for innovative treatments while simultaneously highlighting critical safety concerns regarding the use of such technology.
The newly engineered viruses are a specific type known as bacteriophages, which target bacteria exclusively. They are utilized globally to combat persistent bacterial infections. In laboratory experiments, a combination of these AI-generated bacteriophages demonstrated the ability to eliminate E. coli strains that were resistant to naturally occurring bacteriophages.
Dr. Brian Hie, a chemical engineer at Stanford University, utilized genome language models—akin to the advanced language models powering AI chatbots—to formulate new genetic sequences for bacteriophages. These engineered viruses were synthesized in a lab setting and tested against E. coli in a controlled environment.
The researchers emphasized that the capability to “rapidly design” viral genomes tailored for specific bacterial strains, particularly in overcoming resistance, has the potential to “revolutionize phage therapy” and broaden the biotechnological arsenal, as detailed in their article published in the journal Science.
However, the team acknowledged that their findings bring forth serious biosafety, biocontainment, and biosecurity challenges. They recommended that anyone involved in designing complete genomes should engage with safety and security professionals throughout the development process.
In a related commentary, Prof. Tom Inglesby and Dr. Moritz Hanke from the Johns Hopkins Center for Health Security underscored the urgency of addressing biosafety and biosecurity issues. They pointed out that although this advancement is promising for the life sciences, it urgently raises questions about governance. They noted that while the technology for generating viral genomes through AI now exists, the regulatory framework to manage its safe application is lacking.
The team employed AI models called Evo1 and Evo2 to innovate the viral genomes, having trained these models on genetic material from 2 million bacteriophages. To minimize risks associated with engineering harmful viruses, the AI's training specifically excluded genetic sequences from viruses affecting humans, animals, or plants.
From the thousands of genomes the AI produced, the researchers handpicked nearly 300 for laboratory synthesis. These were introduced to bacteria, which interpreted the genetic instructions and created the new bacteriophages. Although only 16 of these proved viable, the mixture effectively countered resistance in two strains of E. coli.
Despite the compact nature of bacteriophage genomes, Inglesby and Hanke noted that the experiment confirmed the potential of generative AI to yield functional viral genomes. Whether this technique could extend to other viral forms remains uncertain, although the researchers advised against pursuing similar work on pathogens capable of infecting humans, animals, or plants, as such creations could lead to uncontrollable new threats.
Tom Ellis, a professor specializing in synthetic genome engineering at Imperial College London, remarked on the technical achievements of this work but indicated the difficulties of creating more complex genomes. He pointed out that this task involved the simplest and smallest genome to engineer.
Ellis further mentioned that if AI were trained on the genetic data of dangerous pathogens, it could potentially design harmful viruses. However, he suggested that controlling access to genetic information and imposing limits on creating genomes that appear hazardous would greatly assist in mitigating risks. He noted that governments are actively working on such measures.
He concluded by stating, “The perceived threat from fully AI-designed viruses or bacteria is often exaggerated, particularly when you consider how much easier it is to modify existing pathogens to enhance their viral capabilities.”
Dr. Filippa Lentzos, a specialist in science and international security at King’s College London, emphasized the necessity of intervening during the manufacturing of DNA. She advocated for a comprehensive governance approach rather than singling out regulations for AI models. “A multifaceted strategy encompassing safeguards around model development, responsible research assessments, synthesis screening, and established laboratory safety protocols makes more sense,” she suggested.
