Synthetic biology breakthrough: Scientists use AI to create viable bacteriophages from scratch
Researchers have successfully used genome language models to design complete, functional bacteriophage genomes, marking the first generative design of whole living organisms. The breakthrough demonstrates AI’s capacity to access novel evolutionary spaces and create phage cocktails that rapidly overcome bacterial resistance.
Scientists have achieved a milestone in synthetic biology by using artificial intelligence to design complete, functional bacteriophage genomes from scratch. The research, published in bioRxiv on 17 September 2025, represents the first successful generative design of viable whole genomes, opening new possibilities for phage therapy and biotechnology applications.
The international team, led by researchers at Arc Institute and Stanford University, leveraged advanced genome language models called Evo 1 and Evo 2 to generate novel bacteriophage sequences. These AI systems, trained on vast genomic datasets, learned evolutionary constraints that enabled the creation of phage genomes not seen in nature.
Novel AI approach produces functional phages
The researchers focused on designing variants of ΦX174, a well-studied bacteriophage with a compact 5.4-kilobase genome containing 11 genes. Using the phage’s non-pathogenic host, E. coli C, as their target, the team developed a comprehensive computational framework that combined generative modelling with systematic design constraints.
“Many important biological functions arise not from single genes, but from complex interactions encoded by entire genomes,” the authors note. “Here, we report the first generative design of viable bacteriophage genomes.”
The AI models underwent supervised fine-tuning on approximately 15,000 Microviridae genome sequences before generating novel designs. The team then applied stringent filtering criteria including sequence quality control, host specificity requirements, and evolutionary diversity constraints.
From thousands of computationally generated sequences, the researchers synthesised and experimentally tested 302 diverse genome candidates. This systematic screening yielded 16 viable phages with substantial evolutionary novelty, some sharing as little as 63% average amino acid identity with natural proteins.
Structural insights reveal evolutionary innovations
Detailed analysis of the generated phages revealed remarkable structural and functional innovations. Cryo-electron microscopy of one variant, Evo-Φ36, demonstrated successful incorporation of an evolutionarily distant DNA packaging protein from phage G4 – a modification previously thought non-viable in wild-type ΦX174.
“Despite divergent sequence contexts, Evo-Φ36 J preserves a compatible interaction with its capsid, denoting how generative design can uncover novel protein-protein co-evolutionary solutions,” the researchers explain.
The generated phages exhibited diverse mutations including novel gene insertions, extended non-coding regions, gene losses, and substantial protein truncations or elongations. Several contained mutations that could not be attributed to any known natural sequences, highlighting the AI’s capacity to explore uncharted evolutionary territory.
Enhanced fitness and therapeutic potential
Competition assays revealed that several generated phages outperformed their natural template. Three variants – Evo-Φ69, Evo-Φ100, and Evo-Φ111 – consistently ranked among the top performers across multiple fitness competitions, with Evo-Φ69 achieving cumulative fold changes between 16× and 65× compared to ΦX174’s 1.3× to 4.0× range.
Lysis kinetics analysis identified Evo-Φ2483 as exhibiting significantly faster and stronger lytic capabilities than ΦX174, reaching minimum host population density in 135 minutes compared to 180 minutes for the natural phage.
Overcoming bacterial resistance
Perhaps most significantly for therapeutic applications, cocktails of generated phages demonstrated superior ability to overcome bacterial resistance. When challenged with three ΦX174-resistant E. coli strains, the AI-designed phage cocktail successfully inhibited growth within 1-5 passages, whilst ΦX174 alone failed to overcome resistance.
Genomic analysis revealed that successful resistance-overcoming variants arose through recombination and mutation events involving multiple generated phages, with key mutations concentrated on viral surface proteins responsible for host binding.
Broader implications for synthetic biology
The research establishes generative genomics as a powerful approach for accessing novel evolutionary spaces beyond natural selection constraints. The team’s methodology provides a blueprint for designing synthetic biological systems with user-specified properties.
“This work provides a blueprint for the design of diverse synthetic bacteriophages and, more broadly, lays a foundation for the generative design of useful living systems at the genome scale,” the authors conclude.
The computational framework developed encompasses model fine-tuning, prompt engineering, inference-time guidance, and systematic experimental validation. This integrated approach addresses the fundamental challenge of reconciling complex genomic features including coding sequences, regulatory elements, architectural arrangements, and inter-element interactions.
Future directions and safety considerations
The researchers emphasise robust safety frameworks underpinning their work, including computational safeguards within the AI models themselves. The training data deliberately excluded eukaryotic viruses, including human pathogens, whilst experimental protocols followed established biosafety guidelines for bacteriophage research.
Looking ahead, the team envisions applications extending beyond phage therapy to encompass broader biotechnological tools and potentially larger, more complex genomes. The methodology could accelerate development of adaptive antimicrobial strategies and enable systematic exploration of evolutionary design principles.
The research demonstrates that AI-driven generative design can successfully bridge the gap between computational prediction and experimental validation in synthetic biology, offering new avenues for creating functional biological systems with desired properties.
Reference
King, S. H., Driscoll, C. L., Li, D. B., et. al. (2025). Generative design of novel bacteriophages with genome language models. bioRxiv. https://doi.org/10.1101/2025.09.12.675911
