Scientists trained AI on genetic sequences to design viruses not found in nature, yielding viable viruses that can infect bacteria but pose no threat to humans (Carl Zimmer/New York Times)

The research demonstrates that a generative model can produce viable viral genomes that do not exist in nature. By training on libraries of DNA sequences, the AI learned patterns that govern viral structure and function. The output genomes were then synthesized and tested, resulting in viruses capable of infecting bacterial hosts. This shifts the virus discovery process from isolation to computational design.

Adopting this approach requires access to large genetic datasets and expertise in training deep learning models on sequence data. After model generation, researchers must invest in DNA synthesis and laboratory assays to confirm infectivity and safety. The cost includes computational resources for training and inference, as well as synthetic biology workflow expenses. Teams without these capabilities would need to collaborate or outsource parts of the pipeline.

The method currently yields viruses that are limited to bacterial hosts and are engineered to lack human-infectivity traits, so it does not produce threats to humans. If the goal were to target eukaryotic cells or broader host ranges, the same pipeline may not succeed without additional constraints. Furthermore, the reliability of the designs depends on the completeness and bias of the training data; gaps can lead to non-functional or unsafe outputs. Consequently, the technique stops working when the AI proposes genomes that cannot be folded into viable particles or that exhibit unintended activity.