Artificial intelligence has moved beyond predicting individual protein folds to directly generating entire biological genomes. In a roundtable discussion hosted by MIT Technology Review on October 16, 2026, Stanford University bioengineering doctoral researcher Samuel King detailed work conducted with assistant professor Brian Hie and the Arc Institute. The team fine-tuned the genomic foundation model Evo 2 on the bacteriophage phiX174 template to generate complete viral genome blueprints. Out of 285 computationally designed candidates physically synthesized into DNA, 16 successfully manifested as functional viruses capable of infecting and killing E. coli bacteria, achieving an experimental hit rate of approximately 5.6 percent. Several of the artificial strains exhibited replication rates and bacterial lysis speeds superior to the wild-type virus, and combined cocktails of these generated variants completely eliminated three bacterial strains resistant to natural phages.

The demonstration highlights how data scaling in biological foundation models directly translates to functional capability. While the earlier Evo 1 model was trained on 300 billion base pairs primarily sourced from bacteria and phages, Evo 2 expanded more than 26-fold to between 8.8 trillion and 9.3 trillion base pairs covering all domains of life. This massive expansion in training tokens enabled model parameters to capture complex, nonlinear regulatory interactions across entire genomes rather than isolated genetic segments. Brian Hie explained that while bacteria can rapidly develop resistance to a single phage, therapeutic cocktails combining genetically distinct phages make comprehensive resistance far more difficult to acquire, emphasizing the strong utility of AI design for phage therapies targeting antibiotic-resistant superbugs.

Despite the therapeutic promise, the rapid leap in generative biology has intensified biosecurity concerns. The researchers implemented safeguards by restricting their design target to non-pathogenic bacteriophages and strictly filtering human-infecting viral sequences from the pretraining dataset. However, former CDC official Susan Monarez warned that malicious actors could potentially exploit similar frameworks under the guise of legitimate research to design pathogens that evade detection and containment. Existing screening protocols at commercial DNA synthesis providers match orders against known databases of natural pathogens, leaving a structural blind spot for novel, AI-generated synthetic sequences. As global regulators consider classifying computational genome generation under biological weapons frameworks, drug development organizations face rising operational compliance costs alongside synthetic efficiency gains.