Many biological functions emerge from complex interactions encoded across entire genomes rather than from single genes. This study reports the first end-to-end generative design of complete bacteriophage genomes using genome language models, an approach that applies generative artificial intelligence and large language model techniques to viral genome sequences. The authors posit that genome-scale generative design can produce viable viral genomes with actionable properties, such as defined host tropism, and can serve as a blueprint for synthetic biological systems designed at the level of whole genomes.
The research used the well-studied bacteriophage ΦX174 as the design template. Using genome language models, the team generated complete phage genomes with intended host specificity. The approach targets genome-scale patterns and interactions that determine phenotypes such as host range and fitness, moving beyond single-gene edits to generate integrated genome sequences predicted to encode functional phage particles with the desired tropism.
Generated genome sequences were synthesized and experimentally tested in laboratory conditions. The work yielded 16 generated phages that were viable in vitro. These generated phages displayed a range of fitness profiles when assayed under the experimental conditions, indicating functional diversity among AI-designed genomes. The reported experimental validation confirms that genome language models can produce complete viral genomes capable of generating infectious, replication-competent particles.
To verify structural consequences of genome-level design, the investigators used cryo-electron microscopy. Cryo-EM analysis confirmed that at least one generated phage assembles a capsid that incorporates an evolutionarily distant DNA packaging protein. This structural observation provides direct evidence that the generated genome sequence produced nontrivial protein composition and assembly outcomes, consistent with the model-driven design of genome-encoded structural components.
The study tested the functional impact of the generated phages against bacterial hosts, including strains resistant to the parental ΦX174 phage. A cocktail of generated phages rapidly overcame ΦX174-resistant Escherichia coli strains in laboratory assays. This result demonstrates proof-of-concept that collections of AI-designed phages can be combined to address bacterial resistance and suggests a potential path toward phage therapy strategies that adapt to rapidly evolving bacterial pathogens.
The reported results provide a blueprint for designing diverse synthetic bacteriophages and other genome-scale biological systems using generative AI. Key implications include:
Feasibility: Genome language models can generate complete viral genomes that produce viable, structurally validated phage particles.
Diversity and functionality: Designed genomes yield phages with diverse fitness profiles and novel structural features.
Therapeutic potential: Generated-phage cocktails can overcome resistance in target bacterial strains, indicating utility for therapeutic development against evolving pathogens.
Methodological foundation: The integration of large language models with experimental synthesis, functional assays, and structural validation establishes a workflow for future genome-scale design efforts.
The authors frame this work as a foundational demonstration rather than as an immediate clinical intervention; the report presents experimental proof-of-concept for AI-driven genome design and highlights a path toward engineered phage-based interventions.
This work is reported in Science (2026 Aug 6;393[6811]:eaec2657) and is indexed with PMID 42561074 and DOI 10.1126/science.aec2657. The article lists affiliations including Stanford University, Arc Institute, and the Broad Institute, and includes contributions from multiple authors. The abstract and MeSH indexing emphasize topics such as genome language models, large language models, cryo-electron microscopy, capsid proteins, and phage therapy.
Note: The summary above is based solely on the information provided in the source abstract and PubMed record. Detailed methods, quantitative data, and full experimental protocols were not reported in the PubMed abstract and therefore are not included here.