AI Model Trained on DNA Creates 16 Novel Viruses, Raising Both Therapeutic Promise and Biosecurity Alarms
核心洞察
Scientists at Stanford University and Arc Institute (搜索) used AI genome language models Evo 1 and Evo 2 to design entirely new viruses capable of infecting E. coli (搜索) bacteria.
The models, trained on trillions of nucleotides and roughly 15,000 viruses, generated 700,000 candidates; 16 proved viable, with some reproducing faster than the natural Phi X-174 (搜索) virus.
The research, published in Science, could accelerate gene therapy and biotechnology breakthroughs but has drawn urgent warnings from biosecurity experts about potential misuse for biological weapons.
Researchers at Stanford University and Palo Alto's Arc Institute (搜索) have demonstrated that artificial intelligence can design entirely new, functional viruses from scratch—a milestone published in Science that simultaneously opens doors to transformative gene therapies and raises urgent biosecurity concerns.
The team employed genome language models called Evo 1 and Evo 2, which operate on principles analogous to large language models like ChatGPT but are trained on genetic sequences rather than text. These models studied trillions of nucleotides—the fundamental building blocks of DNA—learning what researchers describe as the "grammar" of genetic code.
For the experiment, the scientists further trained the models on approximately 15,000 viruses from the same family as Phi X-174 (搜索), a diminutive virus that exclusively infects Escherichia coli bacteria. The team explicitly constrained the scope to exclude any pathogens capable of infecting humans, other animals, plants, or fungi.
Evo generated roughly 700,000 potential new viral sequences. From this vast pool, the researchers selected the 285 most promising candidates, synthesized their DNA, and inserted the genetic material into bacteria. Ultimately, 16 of Evo's creations yielded working viruses. These AI-designed viruses proved at least as viable as the natural Phi X-174 (搜索), with some demonstrating faster reproductive rates.
Therapeutic Potential and Scientific Promise
The implications for medicine and biotechnology are substantial. AI-designed viruses could enhance gene therapy delivery systems, deepen researchers' understanding of genome architecture, and yield more effective tools for targeting harmful bacteria. In a tightly regulated environment with robust ethical guardrails, the approach represents a potential breakthrough with far-reaching benefits for human health.
Biosecurity Warnings from Experts
The achievement has drawn sharp warnings from biosecurity scholars. Dr. Thomas Inglesby and Dr. Moritz Hanke of the Johns Hopkins Center for Health Security (搜索), writing an accompanying commentary in Science, cautioned that the Stanford research demonstrates AI is now capable of inventing dangerous bioweapons. They called for strict legal frameworks governing future research and argued that using similar generative techniques on pathogens affecting humans, animals, or agricultural crops should be prohibited.
"You could say, 'Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal,'" Dr. Hanke told The New York Times, illustrating the potential for misuse.
Countervailing Perspectives
Not all experts share the same level of alarm. Tom Ellis, a professor of synthetic genome engineering at Imperial College London, told The Guardian that creating a human-impacting bioweapon may not be as straightforward as some fear. He noted that the viruses created by Evo are "literally the smallest and easiest genome to make" and suggested that straightforward restrictions on access to genetic data could substantially mitigate risks.
The Broader Regulatory Landscape
The research arrives amid intensifying debate over biological AI safety. Major AI companies have established the Frontier Model Forum, a nonprofit dedicated in part to researching AI-bio risks, developing safety standards, and reducing the likelihood that frontier models could be exploited for biological misuse. Earlier this year, scientists demonstrated that standard AI chatbots could provide step-by-step guidance on assembling deadly pathogens and deploying them in public spaces, underscoring concerns that such models lower barriers for malicious actors seeking dangerous biological information.
The core tension mirrors that of other dual-use technologies: the same capability that could revolutionize medicine could, in ungoverned hands, enable catastrophic harm. As AI moves from describing biological threats to designing them, the question is no longer whether guardrails are needed, but whether they can be implemented before the technology outpaces them.