Scientists Taught AI to Read DNA and It Designed 16 New Viruses

Artificial intelligence models trained to interpret genetic sequences are moving beyond predicting how DNA behaves. Researchers have now demonstrated that such a model can generate complete viral genomes that become functional bacteriophages after being synthesized and tested in a laboratory.

A study published in Science describes how researchers used the genomic AI models Evo 1 and Evo 2 to design bacteriophages, viruses that infect bacteria rather than people. The team experimentally tested hundreds of generated genomes and identified 16 functional phages capable of infecting strains of Escherichia coli.

The work builds on research that was first released as a bioRxiv preprint in September 2025. Its publication in Science brings the findings into the peer-reviewed literature as researchers increasingly explore whether generative AI can be used not only to analyze biological systems, but also to propose genetic sequences that function when physically constructed.

AI Pushes Beyond DNA Analysis

Evo models genetic information in a way that has parallels with language models trained on text. Instead of learning patterns among words and sentences, the system learns statistical relationships across nucleotide sequences. Those relationships can then be used for tasks ranging from predicting genetic effects to generating new stretches of DNA.

The latest experiment focused on bacteriophage ΦX174, a relatively small virus that infects E. coli. Researchers further trained the Evo models on 14,466 sequences from the Microviridae family of bacteriophages. They then generated candidate genomes intended to resemble ΦX174 closely enough to function while containing substantial genetic differences from naturally occurring sequences.

Researchers did not simply accept the model's output as viable biology. Generated sequences went through computational filtering before selected DNA designs were physically assembled and introduced into laboratory bacteria. A total of 285 designs were experimentally tested, with 16 candidates producing growth inhibition and subsequently being sequence-verified, propagated and examined for fitness and host range. Stanford University separately reported that the team synthesized and tested nearly 300 AI-designed phages before narrowing the group to 16.

AI-Designed Phages Differ From Natural Viruses

The resulting phages were not simply digital copies of ΦX174. Arc Institute said each of the 16 functional genomes contained between 67 and 392 mutations relative to its nearest known natural genome. Thirteen contained mutations that researchers could not identify in known natural sequences, indicating that the model had generated combinations extending beyond sequences already catalogued from nature.

One of the designs, known as Evo-Φ36, incorporated a DNA-packaging protein associated with a more distantly related bacteriophage. Researchers used cryo-electron microscopy to examine how that protein was incorporated into the viral structure. The experiment provided evidence that a genome-scale model could produce a set of coordinated genetic changes that remained compatible with a functioning phage rather than generating isolated DNA components that failed when combined.

The team also examined how the generated viruses responded when bacteria developed resistance. Researchers produced three E. coli strains resistant to ΦX174 and exposed them to mixtures of AI-generated phages. The cocktails overcame resistance in all three strains within one to five experimental passages, while the original ΦX174 phage did not. Those laboratory results suggest a possible research path toward designing more diverse phage populations, although they do not establish that the approach is ready for clinical treatment in humans.

New Applications Bring Biosafety Into Focus

Bacteriophages have long been studied as possible tools against bacterial infections, particularly as scientists search for additional ways to address bacteria that become resistant to existing treatments. Designing phages computationally could eventually give researchers another way to explore candidates instead of relying entirely on viruses discovered in nature. The current study, however, remains an experimental demonstration using laboratory bacterial strains rather than a clinical test of a new therapy.

The work also puts biosafety closer to the center of the debate around generative biological models. Evo 2 was developed by Arc Institute in collaboration with NVIDIA and researchers from Stanford University, UC Berkeley and UC San Francisco. Its training data span trillions of nucleotides across different forms of life, but the developers deliberately excluded genomic sequences from viruses that infect eukaryotic organisms as a safety measure. The peer-reviewed Evo 2 study published in Nature reports that these exclusions weakened the model's performance on human-infecting viruses and that attempts to generate proteins from such viruses produced essentially random results.

Those safeguards do not remove the broader governance questions surrounding increasingly capable biological AI systems. The Evo 2 researchers themselves note that task-specific post-training could potentially circumvent some built-in protections and should be approached cautiously.

The bacteriophage experiment also shows why researchers are pursuing the technology. The model generated genetic designs that survived the transition from computer output to functional biological entities, resulting in 16 laboratory-tested bacteriophages.