US Scientists Create First AI-Engineered Virus

Evolution is constantly writing new biological innovations within genomes. Exploring this vast design space holds the potential to unlock transformative biotechnology functions; however, even the simplest genomes are immensely complex, where a single nucleotide mutation can render an organism non-viable. Consequently, most advances in biological engineering have remained restricted to individual genes or genetic circuits, while whole-genome-scale de novo design has long been deemed an elusive goal.

Recently, a research team led by Brian L. Hie at Stanford University published a landmark study online in Science titled "Generative design of bacteriophages with genome language models."

This study provides the first proof-of-concept that genome language models can generate complete, viable, and functional viral genomes from scratch—much like GPT writes articles—producing synthetic bacteriophages that are distinct from any known natural viruses.

Genome language models are artificial intelligence algorithms trained on massive DNA corpora containing millions of genomes spanning all domains of life. Analogous to how ChatGPT learns human language syntax, the Evo series models learn the evolutionary constraints that shape natural DNA sequences.

Using the natural bacteriophage ΦX174 as a design template, the research team leveraged Evo 1 and Evo 2 to generate thousands of full-length phage genomes with authentic genetic architectures. Nearly 300 of these AI-designed genomes were chemically synthesized and tested under laboratory conditions, ultimately yielding 16 functional, active phages.

Figure 1. A framework for AI-guided bacteriophage genome design. (King S H, et al., 2026)

AI-Written Genomes Unseen in Nature

These active phages demonstrated robust host specificity and diverse fitness characteristics, including competitive infection dynamics. Crucially, the generated phages are distinct from any known natural bacteriophages—harboring de novo mutations, divergent gene and regulatory element compositions, and variable genome lengths.

In one striking instance, a generated phage incorporated a DNA packaging protein sourced from an evolutionarily distant phage family into its capsid structure—a feat equivalent to the AI "borrowing parts from the toolkit of entirely different species" while composing a novel genome.

Overcoming Bacterial Antibiotic Resistance

The researchers evaluated whether these synthetic phages could overcome bacterial resistance—a paramount hurdle in developing phage-based antimicrobial therapies. The results demonstrated that a cocktail of engineered phages rapidly suppressed E. coli strains resistant to wild-type ΦX174, whereas a comparable cocktail of natural ΦX174-like variants failed to do so.

Significance, Potential, and Caution

It is important to objectively note that this study currently focuses on small bacteriophage genomes (~5 kilobases). Scaling this generative approach to larger, more complex genomes—such as bacterial or eukaryotic genomes—will require overcoming substantial computational and synthetic bottlenecks. Moreover, the biosecurity, safety, and ethical implications of AI-generated genomes warrant thorough evaluation.

Nevertheless, this breakthrough establishes a paradigm shift in synthetic genomics. It marks the first time a generative model has captured evolutionary constraints within DNA sequences with sufficient accuracy to yield viable, whole genomes harboring preset traits that diverge significantly from anything observed in nature.

Whether for combating rapidly evolving superbugs, engineering novel phage immunotherapies, or designing larger synthetic genomes in the future, this ability to "write genomes with AI" represents a historic milestone in biological engineering.