Researchers at Stanford University, using two AI models named Evo 1 and Evo 2, have successfully designed and generated complete, functional genomes for bacteriophages—viruses that infect and kill bacteria. This marks the first instance of AI creating an entire living organism's blueprint. The models were trained on phage genomes, similar to how ChatGPT learns from text, using the well-studied ΦX174 bacteriophage as a guide. Out of approximately 300 potential phage genomes designed by the AI, 16 resulted in viable viruses capable of infecting and killing E. coli bacteria in lab settings. Some of these AI-designed phages even performed more effectively than ΦX174, killing bacteria more quickly.
Beyond basic E. coli strains, cocktails of these AI-generated phages were also effective against three E. coli strains that were resistant to ΦX174. This suggests a significant potential for AI in developing phage therapy, a treatment option for antibiotic-resistant bacterial infections. The synthetic viruses were notably diverse, with many qualifying as new species, having less than 95% genetic identity to naturally occurring phages. One AI-designed phage, Evo-Φ69, replicated 16 to 65 times its initial level in host cells within six hours, significantly outperforming ΦX174, which replicated only 1.3 to 4.0 times.
While the immediate applications point to medical advancements and microbial manufacturing, experts express caution about the broader implications. Craig Venter, known for his work in synthetic DNA, warned against viral enhancement research, especially concerning human pathogens. Yale Law School student Tal Feldman and Georgia Tech researcher Jonathan Feldman highlighted the risk of bad actors using AI to create novel bioweapons, emphasizing that the world is unprepared for the rapid pace at which AI can design such viruses. The Stanford team intentionally avoided training their AI on human pathogens to mitigate these risks, focusing instead on bacteriophages, which do not infect humans.
