The genome language model Evo 2 has designed novel functional genomes of bacteriophage ΦX174. About 300 of the AI-generated designs were chemically synthesised and tested in the laboratory. Sixteen synthesised phages were viable and were found to be effective against the resistant Escherichia coli C bacteria, suggesting that phage therapy could be an option for dealing with bacterial resistance. As a milestone in science, this is the first time a language model has designed biological function at the whole-genome level.
Evolution has been at work for about 3 billion years designing biological systems in nature through random mutation and selection. People have used artificial selection for quite some time to shape biological features of plants and animals. Genetic engineering is an established tool now for designing biological systems, courtesy of advances in molecular biology in the past few decades. This requires predicting structural details of a functional and viable design, which is a complex and challenging task. Vast computing power of modern LLMs suitably trained with relevant evolutionary and genetic data has come in handy in augmenting the search of the functional biological designs.
Of late, generative AI has found application in designing complex biological systems at the scale of individual genes and proteins. In 2025, researchers showed that a language model trained at scale on evolutionary data can generate novel functional proteins. They presented ESM3, a protein language model which generates protein structure and amino acid sequences when prompted. Protein sequences determine three-dimensional structure and cellular function of a protein.
Genome language models are trained on large corpora of DNA comprising millions of genomes from all domains of life. As a result, these models learn the evolutionary constraints that shape DNA sequences in nature. Can a genome language model generate entire functional genomes?
In a recent study reported on 6 August 2026, researchers have successfully used the genome language models, Evo 1 and Evo 2, to generate complete genomes of bacteriophage ΦX174 with realistic genetic architectures and specificity for the bacterial host, Escherichia coli C. They used natural phage ΦX174 as a design template. Thousands of the generated genomes were evaluated, of which about 300 were chemically synthesised and tested in the laboratory. Sixteen phages were found viable. The viable phages were novel and rapidly overcame ΦX174-resistant E. coli strains when tested for this ability.
This study elevates genome language models to the next level as it has shown that these models can capture evolutionary constraints in DNA sequences with enough fidelity to produce complete functional genomes. This also lays out a foundational path for bacteriophage therapies against resistant bacteria.
***
References:
- Hayes T. et al. 2025. Simulating 500 million years of evolution with a language model. Science. 16 January 2025. Vol 387, Issue 6736 pp. 850-858. DOI: https://doi.org/10.1126/science.ads0018
- King S. H. et al. 2026. Generative design of bacteriophages with genome language models. Science. 6 August 2026. Vol 393, Issue 6811. DOI: https://doi.org/10.1126/science.aec2657
- Inglesby T.V. and Hanke M.S., 2026. AI-designed viral genomes. Science. 6 August 2026. Vol 393, Issue 6811 pp. 563-564. DOI: https://doi.org/10.1126/science.aej8512
- AI designs a novel E. Coli killer. Stanford Reports. 6 August 2026. Available at https://news.stanford.edu/stories/2026/08/evo-2-ai-tool-e-coli-killer-bacteriophages
***
Related articles:
- Brain-Computer Interfaces (BCI): Towards Humans’ Merger with AI (18 March 2026)
- SARAH: WHO’s first generative AI-based Tool for Health Promotion (2 April 2024)
- Generative Artificial Intelligence (AI): WHO issues new Guidance on governance of LMMs (19 January 2024)
- Artificial Intelligence (AI) Systems Conduct Research in Chemistry Autonomously (26 December 2023)
- Artificial Intelligence (AI) for Fast and Efficient Medical Diagnosis (15 April 2018)
***
