Bacteriophages are viruses that infect and kill bacteria. Scientists have long studied them as a possible alternative to traditional antibiotics. Chemical engineer Brian Hie and graduate student Samuel King focused on a simple bacteriophage called ΦX174, whose genome contains fewer than 6,000 base pairs.
Hie developed Evo 2, a generative artificial intelligence model that learns patterns from large collections of DNA and then suggests entirely new DNA sequences. Given only a short fragment of ΦX174 DNA, the model produced thousands of complete candidate genomes for new bacteriophages that target the bacterium E. coli. Researchers chemically built nearly 300 of these genomes and introduced them into bacteria. Laboratory tests showed that 16 of them formed functional viruses able to infect and destroy E. coli. Some of these artificial bacteriophages even performed better than the original natural virus.
Fighting bacterial resistance with diverse mixtures
Bacteria often evolve resistance to single treatments, making antibiotics less effective over time. A mixture of genetically different bacteriophages can make it harder for bacteria to develop full resistance. Tests confirmed that a combination of the 16 new viruses quickly overcame E. coli strains that had already become immune to the natural ΦX174. Similar approaches could one day target other dangerous bacteria that cause hard-to-treat infections.
The process of designing and building entire genomes remains technically demanding and costly. King created a computational framework that evaluated the model’s suggestions against useful biological traits, allowing researchers to select only the most promising candidates for laboratory testing. Evo 2 is freely available so others can use it to explore new genome designs. While the open release has prompted discussion of safety, the researchers note that existing natural pathogens currently pose greater practical risks and that tools of this kind can also strengthen defenses against biological threats. Future work aims to increase the novelty and control of the generated sequences and to apply the approach to longer, more complex DNA.
This research is published in Science.