AI-designed virus genomes raise concerns about future human pathogens
Artificial intelligence models trained on genome sequences have been used for the first time to generate complete genetic blueprints of viruses that infect bacteria. Researchers at Stanford University stress that a similar approach could eventually be applied to design pathogens dangerous to vertebrates, and suggest that safeguards should be considered now.
Most previous AI work in biology focused on designing proteins, the molecules that carry out most cellular functions. Since the genetic code acts as an intermediate layer between DNA and proteins, the potential of models trained on nucleotide sequences was initially unclear. However, experiments showed that such systems can produce DNA encoding functional proteins in bacteria and mimic gene structures found in complex cells. Now the same technology has been used to generate entire virus genomes.
All the viruses created by the models are closely related to known bacteriophages. Yet they possess distinct features that would be difficult to achieve through natural evolution. This, according to the authors, highlights the need to assess potential risks in advance: in the future, similar AI could be used to engineer viruses targeting mammals.
Large genome models work on the same principle as large language models, which predict the next fragment of text based on vast datasets. In DNA, the alphabet is limited to four letters — A, T, C, and G. However, the task is complicated by the fact that genomes contain regions where the next nucleotide is critical, interspersed with sequences where any substitution would be irrelevant. The model must account for biological context even when humans have not yet fully deciphered it.
Practical tests showed that when prompted with a fragment of a bacterial gene cluster carrying related functions, the model outputs DNA sequences encoding proteins with similar roles. Some of these may be functional variants unlike any protein known to science. At the same time, the limits of human knowledge remain a constraint: when prompted with a sequence from a complex cell, the model responds with a string of bases resembling genes, but their actual function requires further study.
