Breakthrough study uses AI to produce 16 new bacteria-killing viruses
Scientists have used artificial intelligence to create blueprints for 16 viruses that do not occur in nature but are capable of killing hazardous E. coli bacteria. The breakthrough could lead to new treatments for antibiotic-resistant illnesses, but it also raises new biosecurity concerns, researchers say.
The viruses, called bacteriophages, were designed from scratch by AI, not replicated from nature.
The work by a Stanford research team, published in Science, used an AI model called Evo 2, which was trained to interpret genetic code much as large language models learn human language.
Instead of predicting words, Evo 2 predicts DNA sequences. Researchers fed it a tiny piece of genetic material from a naturally occurring virus that infects E. coli, then challenged it to create entirely new genomes to do the same job.
The AI created almost 700,000 possible virus genomes. The scientists whittled them down to about 300 choices, synthesised them in the laboratory and checked if they worked. Sixteen worked inside E. coli cells and killed the bacterium, some even better than the original virus they were based on.
Breakthrough fraught with risk
The achievement has also reopened debate about the dangers of ever more powerful artificial intelligence systems in biology. The Stanford team did not train Evo 2 on human-infecting viruses, only on bacteriophages, but biosecurity experts say the research points out the need for stronger safeguards as AI becomes able to design increasingly complex organisms.
The researchers claim that standard artificial intelligence can also enhance safety by enabling scientists to respond more swiftly to developing diseases and by allowing protections to be embedded directly into AI technologies.
Antibiotic resistance is now considered one of the largest global health problems as germs increasingly develop ways to survive existing antibiotics. Bacteriophages, which naturally infect and kill certain bacteria, are increasingly being considered as a potential alternative.
