Researchers have created new viruses using AI. The hope is that the tech can be used to create targeted phages ** — that take out problematic bacteria — selectively killing an antibiotic-resistant strain infecting a person, for instance. However, “there is no sugarcoating the risks,” Tal and Jonathan Feldman wrote in The Washington Post
Researchers have created new viruses using AI. The hope is that the tech can be used to create targeted phages ** — that take out problematic bacteria — selectively killing an antibiotic-resistant strain infecting a person, for instance. However, “there is no sugarcoating the risks,” Tal and Jonathan Feldman wrote in The Washington Post
It is a double edge sword say the Feldmans. “The same techniques could just as easily be used to create viruses lethal to humans — turning a laboratory breakthrough into a global security threat.”
The concern isn’t just for potential threat from novel viruses. “The next step is AI-generated life,” Brian Hie, a co-author on the preprint reporting the new genomes, told Nature. “This is the first time AI systems are able to write coherent genome-scale sequences,” he says.
To generate new viral genomes, the researchers fed the real genome a simple bacteria-killing virus to two AI models — Evo 1 and Evo 2 — which analyze and generate genetic sequences, following training on two million similar phage genomes.
The team told the models to create genomes for viruses capable of killing antibiotic-resistant Escherichia coli. Of the thousands of genomes the models produced, the team zeroed in on just over 300 to synthesize and infect bacteria with.
Of those, 16 infected E. coli selectively; certain combinations of them were able to wipe out three strains of the bacterium that ΦX174 was powerless against.
The models were not ready to make whole genomes by themselves — the team needed to give them a template for guidance and sift through the many ”not-so-stellar genomes” they designed.
After all, just as with large language models, Evo 1 and Evo 2 sometimes hallucinate, Hie told IFLScience.
“In this study, our approach to biosafety and biocontainment leveraged computational safeguards inherent to our models, experimental design choices and established laboratory containment protocols,” the authors explained.
The models weren’t trained on viruses that can infect people, for example. But the Feldmans wonder: “What’s to stop others from using open data on human pathogens to build their own models?”
It’s imperative that new ways of detecting potential dangers are developed, Bruce Wittmann and colleagues wrote in the most recent issue of Science.
While their work focused on toxic proteins, their point stands when it comes to viruses and living organisms. “In the long term, sequence-based biosecurity screening alone is unlikely to remain sufficient,” they wrote.
“Accordingly, although efforts to improve sequence-based hazard detection software should continue, we must also pursue the development of new approaches.”
**(Bacteriophages are viruses that infect and replicate only in bacterial cells known)
Read the SCIENCE PAPER