Researchers at Stanford University have demonstrated that large genome models can be used to design new viruses that infect bacteria. The models, named Evo 1 and Evo 2, were trained on DNA sequences and used to generate viral genomes, including those of bacteria-infecting viruses. The study highlights the potential of these models to create functional viral sequences that can inhibit bacterial growth. The models were tested on the ΦX174 virus, a well-studied bacteriophage, and produced a range of outputs that were evaluated for viability. The findings suggest that these models could be used to design new viruses with specific functions, raising concerns about their potential misuse.
The study focused on the ΦX174 virus, which has a simple genome and a well-characterized infection cycle. Evo 1 and Evo 2 were trained on DNA sequences from bacteriophages and fine-tuned with sequences specific to the Microviridae family. The researchers tested various prompts to generate viral sequences, finding that prompts with four to nine bases of the ΦX174 start sequence worked best. They discarded outputs that did not meet specific criteria, such as having a spike protein with less than 60 percent similarity to the original or being too long or too short. This process left them with 302 potential outputs, 285 of which were chemically synthesized and tested.
The researchers inserted the synthesized sequences into bacteria and observed the effects. Most sequences had no effect, but 16 inhibited E. coli growth, indicating they functioned as viruses. Nine of these were initial outputs from the AI, while the remaining seven had acquired mutations after insertion. The study also found that sequences most similar to ΦX174 had a higher viability rate, suggesting that small changes can significantly impact functionality. The results highlight both the potential and the risks of using large genome models to design new viruses.
Source: arstechnica
