Stanford researchers have moved generative biology from proposed DNA sequences into living laboratory tests. Using the Evo 2 model, the team generated thousands of candidate bacteriophage genomes, synthesised nearly 300 and identified 16 with particularly strong activity against E. coli.
The experiment centres on ΦX174, a compact phage with fewer than 6,000 base pairs. Some generated variants reportedly showed higher laboratory fitness than the native virus. The result is notable because Evo 2 produced complete genomes rather than suggesting isolated edits—but computational screening, chemical synthesis and biological assays remained essential at every step.
Genome generation moves from sequence to synthesis
Evo 2 received a short starting fragment and generated an entire ΦX174 genome in one left-to-right pass. The researchers did not manually complete the proposed sequence, according to project leader Brian Hie.
A small phage offers a manageable proving ground compared with the roughly three billion base pairs in a human genome. Even so, interpreting a sequence of approximately 5,400 characters gene by gene is difficult, and a syntactically plausible genome may not create a viable virus.
Graduate researcher Samuel King developed a computational framework to narrow the candidate pool before synthesis. It evaluated traits derived from ΦX174 and related phages, then selected designs judged most promising. That filtering reduced the number of expensive physical experiments.
The workflow defines what the model did and did not achieve. Evo 2 expanded the design space, but it did not predict viability with enough certainty to remove laboratory selection. Thousands of generated genomes became hundreds of synthesised candidates and ultimately a small group with strong measured activity.
Screening and cocktails turn candidates into evidence
Bacteria can evolve resistance to a single phage, so the researchers combined the 16 selected designs. Stanford reports that this mixture rapidly overcame resistance in E. coli that was immune to native ΦX174.
Genetic diversity may make it harder for bacteria to evade every member of a cocktail, but the source does not establish clinical effectiveness. Laboratory killing activity differs from safety, dosing, stability and performance in animals or humans. The work is an experimental design result, not an approved treatment.
The team says similar methods could target pathogens such as methicillin-resistant Staphylococcus aureus or Pseudomonas aeruginosa. Those are research directions rather than demonstrated outcomes. Longer genomes, different host biology and tighter control requirements could make future targets substantially harder.
Better evidence will require independent replication, detailed comparison with conventional phage-engineering methods and tests of how resistance develops over time. Researchers will also need to determine which generated changes produced the observed fitness improvements rather than treating the model as an opaque source of candidates.
Open access expands both opportunity and oversight
Evo 2 is available as open-source software, allowing researchers to download it and explore genome design. Open access can accelerate validation, adaptation and scrutiny across laboratories that work on therapeutics, engineered microbes and biological manufacturing.
It also raises dual-use concerns. Hie acknowledged that malicious actors could modify tools, while arguing that existing pathogens are currently easier to obtain and reproduce. That is his assessment; the balance can change as models become more capable, controllable and accessible.
Responsible use should combine access controls for sensitive workflows, screening of proposed sequences, institutional biosafety review and transparent reporting of limitations. Safeguards must cover the complete pipeline because a digital sequence only becomes a biological risk when synthesis and experimentation follow.
Stanford plans to extend the research to longer, more complex DNA and possibly small bacterial genomes that could produce chemicals, medicines or fuels. The open technical questions identified by the team—greater novelty and greater control—are also governance questions. More novelty broadens discovery, while more control can increase both utility and misuse potential.
Evo 2 demonstrates that a generative model can contribute viable whole-genome designs. The achievement belongs to the integrated system: model, computational filter, DNA synthesis and laboratory validation. In biological engineering, generation opens the search; experimental evidence still decides what works.
