Scientists Create First Viruses Designed by AI
Stanford and Arc Institute researchers produced 16 viable bacteriophages from AI-designed genomes, pointing to new therapies while raising biosecurity concerns.
Topics
News
- Temasek, Seraphim Back Pixxel in $100 Million Round
- UN Rights Chief Warns Advanced AI Could Threaten Humanity
- Tata-owned JLR to Cut About 4,000 Bobs Amid Tariff, China Pressure
- Anthropic Walks Away From $6 Billion Decart Deal
- Hyundai India Targets 20% Women in Executive Workforce by 2030
- Security Breach Drains $320 Million in Bitcoin From Liquid Network
Researchers at Stanford University and the Arc Institute have used generative artificial intelligence to design complete viral genomes that, after being synthesized and tested in a laboratory, produced 16 viable bacteriophages capable of infecting and killing E. coli.
The study, published in Science, is the first peer-reviewed demonstration that generative models can design genome-scale sequences that function as viruses, extending AI-assisted biology beyond individual proteins, antibiotics and other biological components.
The researchers used genome language models called Evo 1 and Evo 2, which learn patterns in genetic sequences much as conventional language models learn relationships between words. Built on broad genomic training data, the models were further trained to generate genomes resembling ΦX174, a well-understood bacteriophage that infects E. coli.
Thousands of designs passed through computational screening before the researchers shortlisted 302 candidates. They synthesized and experimentally tested 285, of which 16 produced viable phages that could replicate inside bacterial cells and destroy their hosts.
“This is a next step in the complexity that’s designable by generative AI,” Brian Hie, an assistant professor at Stanford University, told the BBC.
Phage therapy uses viruses that attack bacteria as an alternative or supplement to antibiotics. Although the approach has existed for decades, interest has increased as antimicrobial resistance reduces the effectiveness of conventional drugs.
Bacteriophages target bacteria rather than human cells, and the researchers worked with nonpathogenic laboratory strains of E. coli. The findings also do not establish that the models could design arbitrary viruses or pathogens affecting people, animals or plants.
The ability to generate complete viral genomes, however, also broadens concerns over how biological AI systems should be tested and controlled.
Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security said in a commentary accompanying the research in Science that the work raises “urgent biosafety and biosecurity questions.”
They argued that the policy challenge is moving beyond whether AI will become capable of designing viral genomes to determining how those capabilities can be developed without enabling harmful applications.
The Stanford and Arc Institute team excluded viruses that infect humans, animals and plants from the relevant training data, confined the work to bacteriophages and used established laboratory controls. The researchers also called for scientists attempting whole-genome design to consult biosafety and security specialists throughout their projects.
The genomes involved were also considerably simpler than those of living organisms. The phages contained about 5,400 genetic base pairs, compared with roughly 500,000 in some of the smallest cellular genomes and about three billion in humans.
