The AI That Learned to Create Viruses from Scratch: A Historic Breakthrough or a Recipe for Biological Disaster?
A groundbreaking study by Stanford University and the Arc research institute: an AI model named Evo, trained on trillions of DNA sequences, designed hundreds of thousands of possible genomes, of which 285 sequences were built and 16 became live, active viruses in the lab. The new viruses, bacteriophages that attack only bacteria, showed high efficacy, and some multiplied faster than the natural virus. Alongside the excitement, biosecurity experts from Johns Hopkins University warn that control and regulation mechanisms have not yet been created.
In a groundbreaking study, researchers from Stanford University and the Arc research institute in California used an advanced AI model called Evo, trained on trillions of DNA sequences. The model, similar to ChatGPT, learns to identify patterns in the language of DNA, which consists of four letters: A, C, G, and T. To ensure safety, the researchers focused the model on the bacteriophage family, viruses that attack only bacteria and cannot infect humans. The model was trained on the Phi X-174 virus and about 15,000 of its relatives. After training, the system generated about 700,000 potential designs, of which 285 sequences were built in the lab. Upon introducing the DNA into E. coli bacteria, 16 of the genomes became live, active viruses. Some multiplied at a faster rate than the natural virus and exhibited characteristics evolutionarily distant from the original. Alongside the excitement, biosecurity experts from Johns Hopkins University warn that the ability to assemble viral genomes exists, but control and regulation mechanisms have not yet been created.