IA · 7 August 2026 · 5 min read
From Prompts to DNA: AI Generates 16 New Viruses and Kicks Off the Genomic Model Era
In brief: A research team from Stanford University and the Arc Institute has demonstrated the ability of AI genomic foundation models to design functional bacteriophages from scratch. By generating and chemically synthesizing 300 candidates, researchers created 16 viable bacterial viruses, offering promising tools to fight antibiotic resistance while raising urgent biosecurity questions.
by Team Mocchi's
A Shift from Bio-Informatics to Genetic Synthesis
Until recently, the application of artificial intelligence in the life sciences focused primarily on designing individual proteins and predicting their three-dimensional structures. However, computational biology has reached a major conceptual milestone: moving from protein models to Large Genome Models (LGMs) capable of understanding and manipulating an organism's entire genetic code.
A team of researchers from Stanford University and the Arc Institute has demonstrated that a generative model can design complex, fully functional viral genomes from scratch. As reported by WIRED, scientists successfully synthesized 16 novel bacteriophage viruses in the laboratory—organisms that do not exist in nature, designed specifically to infect and neutralize bacterial strains like Escherichia coli.
From Words to the Code of Life: How Evo 1 and Evo 2 Work
The architecture underlying this breakthrough relies on the foundational models Evo 1 and Evo 2. Much like large language models operate on text, these algorithms are trained not on human words or grammar, but on millions of genomic sequences from across all domains of life: plants, animals, microbes, and viruses.
The core challenge lies in the fact that the genetic code, composed of the four nucleotide bases A, T, C, and G, alternates between highly constrained regions and variable segments. As detailed by Ars Technica, the model learns the intricate functional organization of genes and regulatory sequences without developers having to explicitly hardcode biological rules. Using the structure of the PhiX174 bacteriophage as a reference blueprint, the AI generated thousands of genetic variations while maintaining the functional capacity to correctly assemble and replicate the virus within a host cell.
From Algorithm to Petri Dish: 16 Functional Viruses
To validate the biological viability of the AI's generated genomes, the researchers selected 300 candidate sequences from the thousands produced by the model. These sequences were chemically synthesized molecule by molecule in the lab and introduced into E. coli bacterial cultures.
The experimental results confirmed the approach: 16 of the 300 synthesized genomes gave rise to fully viable bacteriophages. These AI-designed viruses not only successfully bound to bacterial membranes and injected their genetic material, but completed their full replication cycle, causing bacterial cell lysis. Even though their DNA sequences differed significantly from any naturally occurring phages, their functional architecture operated flawlessly.
The Dual-Use Dilemma: Phage Therapy vs. Biosecurity
The study's findings open up transformative possibilities for medicine and biotechnology. The ability to custom-design bacteriophages provides a strategic new weapon against antibiotic-resistant superbugs, paving the way for targeted biotherapies against complex bacterial infections.
At the same time, demonstrating that AI can engineer autonomous viral genomes raises urgent biosecurity questions. To mitigate immediate risks, Stanford researchers intentionally excluded vertebrate-infecting viruses from the Evo training datasets. Nevertheless, the international scientific community stresses the pressing need to update regulatory frameworks and DNA synthesis screening protocols before the widespread availability of such models enables dual-use applications or the unintentional creation of human pathogens.
Mocchi's take
Extending Transformer architectures and foundation models to genomic code marks a pivotal shift for the applied artificial intelligence landscape. For tech organizations and research labs building advanced AI solutions, this milestone shows that generative models offer immense value when applied to complex, highly structured domain-specific systems beyond text and media generation. Combining generative AI with experimental validation demands strict data governance and robust 'security-by-design' infrastructure. European software and bio-tech ecosystems must lead the charge in pairing this rapid technological innovation with rigorous ethical standards and compliance frameworks.