Scientists at Stanford University in the United States have designed complete, functional genomes for bacteriophage viruses using generative artificial intelligence models. The study, published in the journal Science, details how language models trained on genetic sequences were able to design viable viral genetic material from scratch for viruses that infect bacteria.
The research marks an advance for synthetic biology and for the development of personalized therapies against bacterial infections, according to the study.
Genomic Language Modeling
Unlike AI models built for human text, these systems work as genomic language models, treating nucleotides, the chemical letters of DNA known as A, C, G and T, as words and sentences. The researchers fine-tuned the model using a database made up exclusively of public genomes from viruses in the Microviridae family.
From that dataset, the AI learned the structural rules needed to design new genomes ranging from 4,000 to 6,000 DNA base pairs. Bacteriophages, the class of virus involved, infect and replicate inside bacteria rather than human or animal cells.
Safety and Biosecurity Filters
To ensure the experiment's biosafety, the AI model was trained strictly on bacteriophages, viruses that are unable to infect humans or other eukaryotic organisms.
The Stanford researchers also applied computational design rules known as tropism filters. These parameters ensure the designed viruses attack only the host bacterium chosen for the laboratory test, Escherichia coli C. Data from viruses that infect humans was also preemptively excluded from the model's training to avoid the risk of creating harmful pathogens.
Results and Biological Viability
The sequences generated by the AI were sent to commercial DNA synthesis companies. In the laboratory, the synthetic DNA fragments were assembled and introduced into recipient bacteria to test whether they would produce active viral particles capable of multiplying.
Plating assays and analyses of bacterial growth curves confirmed that the AI-designed genomes were biologically viable, and that the AI-generated viruses matched the predicted models with precision.
Implications for Medicine and Research
The results show that genomic language models can accelerate the design of tailor-made biological agents, offering alternatives for phage therapy against antibiotic-resistant bacteria.
Experts cited in the study said the maturing of generative biology will require constant updates to biosafety protocols, DNA synthesis screening and governance of scientific data.
