On the 6th, American scientists released a new research finding, announcing that they have used artificial intelligence (AI) to design new viruses that do not exist in nature, sparking widespread attention.
A research team led by Stanford University in the United States published a paper in the journal Science, stating that they used AI to design brand new, fully functional bacteriophages that can replicate in the laboratory.
The research team noted that they designed these viruses using a genomic language model. While ordinary AI large language models are trained on vast amounts of text data, a genomic language model is trained on large quantities of natural genomic DNA sequences, learning to recognize DNA structural patterns that have formed through long-term evolution in nature, and then generates new genomes accordingly.
In this study, the team used two genomic language models, Evo 1 and Evo 2, to generate a batch of candidate genomes using the natural phage Phi X-174 as a template. About 300 of these were synthesized artificially and tested in the lab. Ultimately, 16 strains of active bacteriophages were obtained. These bacteriophages have genomes with different DNA sequences, structures, and adaptability.
Bacteriophages are viruses that attack bacteria. The research team said they chose bacteriophages as their research subject because their genomes are small, easy to validate experimentally, and they are widely used in molecular biology, microbial engineering, and medical fields.
The researchers pointed out that these AI-designed bacteriophages differ from any known natural phages and possess new mutations, differentiated genes and regulatory elements, as well as genome lengths. Experiments showed that preparations made up of these novel phages could quickly inhibit a strain of E. coli that had already become resistant to natural bacteriophages.
The research team therefore believes that AI models can design bacteriophage genomes completely different from those found in nature and that possess predefined characteristics.
This study has raised concerns about whether AI could be used to design and create dangerous pathogens.
According to the research team, to reduce risk, they deliberately excluded data on viruses that can infect humans or other animals and plants from AI training, so the resulting models cannot generate viral genomes that threaten humans. The 16 phages created in this study pose no threat to humans.
According to the New York Times, the researchers chose the natural phage Phi X-174 as their design template partly because scientists have studied this virus for nearly a century and understand it extremely well, and partly because Phi X-174 only infects E. coli, meaning similar viruses are relatively safe.
However, the research team also acknowledged that the research "raises important issues at the levels of biosafety, biocontainment, and biosecurity," calling on others carrying out whole-genome design research "to consult professionals in biosafety and biosecurity throughout the entire project cycle."
In a contemporaneous commentary published in the journal Science, public health expert Inglesby and biosecurity expert Hanke from the Johns Hopkins Center for Health Security pointed out that today the question is no longer "whether AI can generate viral genomes" but rather how to ensure related applications "do not pose serious harm."
They called for no development of any viruses with potential pathogenic capabilities and for strengthened oversight through legislation.