Scientists from Stanford University have used artificial intelligence for the first time to design complete viral genomes. According to a study published in the journal Science, AI-generated viruses have successfully multiplied under controlled laboratory conditions. To solve this problem, the researchers used generative Evo1 and Evo2 models, which were trained on extensive arrays of genetic codes of viruses, bacteria, plants and humans. The principle of operation of these algorithms is similar to the functioning of large language models, but instead of predicting text sequences, they predict DNA structures.

During the experiment, the models were adapted to generate bacteriophages, viruses that selectively infect certain types of bacteria. From thousands of generated variants, the researchers selected 302 of the most promising sequences and synthesized them in the laboratory. The test results showed that 16 of them were effective against the bacterium Escherichia coli (E. coli). In some cases, the created viruses have successfully overcome the natural defense mechanisms of bacterial cells. It is noted that the obtained genetic sequences have structural patterns that are not found in the natural environment.

The publication sparked a discussion in the scientific community regarding biosafety issues. In an accompanying commentary to an article in the journal Science, experts from the Johns Hopkins University Center for Health and Safety, Thomas Inglesby and Moritz Hanke, pointed to urgent issues in this area. They warned about the potential risks of using genomic language models to modify pathogens in order to increase their contagiousness or lethality. In response, the authors of the study emphasized that viruses infecting complex organisms were intentionally excluded from the training sample, and all experiments were conducted exclusively with bacteriophages in a secure laboratory.