Researchers have used artificial intelligence to design complete bacteriophage genomes from scratch, then synthesized and tested the resulting viruses against bacteria that had evolved resistance to a naturally occurring phage. The study, published in Science, represents a significant advance in generative genomics, a field seeking to use computational models to create functional biological systems rather than modifying one gene or a small genetic circuit at a time. The work also highlights the growing difficulty of separating scientific opportunity from biosafety and biosecurity risk as DNA design and synthesis become increasingly accessible.
The team, led by Samuel King, focused on ΦX174, a small and extensively studied bacteriophage that infects Escherichia coli. Bacteriophages, or phages, are viruses that reproduce inside bacteria and destroy their host cells during infection. Because they can target specific bacterial strains, phages have attracted renewed interest as potential treatments for infections that no longer respond to antibiotics. Their clinical use, however, is often limited by the rapid evolution of bacterial resistance. Designing phages with altered genetic and structural properties could eventually provide a way to broaden or extend the usefulness of phage-based therapies.
Creating a functional viral genome is far more complicated than assembling a list of genes. A genome must contain protein-coding sequences, regulatory elements, overlapping regions, signals controlling the timing and level of gene expression, and structural instructions that allow the resulting virus to package its genetic material and infect a host. These components interact in ways that are not always apparent from their individual sequences. A change that appears harmless in one region can disrupt the folding, expression, replication, or assembly of the entire virus. For that reason, most genome engineering has historically relied on modifying known natural templates rather than designing whole genomes independently.
King and colleagues addressed this challenge by combining genomic language models with computational biology and laboratory screening. Their approach used the Evo family of models, which had previously been developed to learn patterns across biological sequences. Like language models trained on text, genomic language models process sequences as ordered information and attempt to learn which combinations of nucleotides are plausible in a biological context. In this study, the model was used not simply to predict the effect of a single mutation, but to help generate candidate phage genomes containing coordinated changes across the genome.
The researchers computationally designed hundreds of candidate ΦX174-like genomes. These candidates were then synthesized and tested experimentally to determine whether they could produce infectious, self-replicating phages. Most designs did not meet that demanding standard, underscoring the difficulty of whole-genome generation even when the starting biological system is well characterized. Sixteen candidates were identified as functional. Their sequences and predicted or observed structures differed substantially from one another, indicating that more than one genomic solution can support a working phage.
Several of the engineered viruses performed comparably to naturally occurring relatives in laboratory tests, according to the researchers. This result is important because it suggests that generative models can produce genomes that preserve the many coordinated functions required for infection, replication, assembly, and release. The achievement is not equivalent to designing any virus on demand; rather, it demonstrates that a model-guided process can search a vast sequence space and identify a limited number of viable designs through experimental validation.
The study also examined whether the engineered phages could address resistance to ΦX174-like viruses. Bacterial resistance can arise through changes in surface receptors used by phages to attach to cells, alterations in intracellular defenses, or other mechanisms that block viral replication. The researchers report that combinations of the newly generated phages were able to overcome resistance in two E. coli strains that resisted ΦX174-like phages. Such combinations may be valuable because using multiple phages with different infection properties can make it more difficult for bacteria to escape treatment through a single mutation.
The findings nevertheless come with important limitations. Laboratory activity against selected bacterial strains does not establish therapeutic effectiveness in animals or humans, where immune responses, tissue environments, microbiomes, and pharmacological constraints can alter phage behavior. Nor does the study show that AI-designed genomes are reliably predictable before synthesis. The large gap between the number of proposed genomes and the number that functioned demonstrates that experimental screening remains essential. Future work will need to determine how well these methods generalize to larger and more complex phages, different bacterial hosts, and clinical settings.
The ability to generate complete viral genomes also raises concerns that extend beyond phage therapy. In a related Science Perspective, Thomas Inglesby and Moritz Hanke note that the authors address biosafety and biosecurity more deliberately than many developers of powerful biological AI systems. King and colleagues argue that whole-genome design projects should involve safety and security specialists throughout their development, from model construction and sequence generation to synthesis and laboratory testing. They also suggest that existing biological safety frameworks could be adapted to generative genomics, while model-level safeguards, including the exclusion of sensitive viral sequences from training data, might provide an additional layer of protection.
As sequencing technologies continue to make genomes easier to read and DNA synthesis makes them easier to write, the central challenge will be governing the transition from computational possibility to biological capability. The new study shows that AI can help identify functional designs in a highly complex viral system, while also revealing how much uncertainty remains between a digital sequence and a working organism. “The question is no longer whether generative viral genome design will exist,” Inglesby and Hanke write. “It is whether society can build oversight that allows its benefits to unfold while preventing it from enabling serious harm.”
Subject of Research: AI-guided generative design of functional bacteriophage genomes and their potential to overcome bacterial resistance.
Article Title: Generative design of bacteriophages with genome language models
News Publication Date: 6-Aug-2026
Web References: https://doi.org/10.1126/science.aec2657
References: King and colleagues, “Generative design of bacteriophages with genome language models,” Science.
Keywords
Generative genomics, artificial intelligence, genome language models, bacteriophages, phage therapy, ΦX174, Escherichia coli, bacterial resistance, synthetic biology, biosafety, biosecurity
