Evo 2 AI Designs 16 Viable Bacteriophages in Lab Test
Researchers at Stanford University have used the large genome models Evo 1 and Evo 2 to design working bacteriophages—viruses that infect bacteria. The AI-generated viruses are genetically distinct from the original ΦX174, with some carrying over 50 amino acid changes, yet they still function. The work, published in Science, demonstrates both the potential and the risks of using large language models for genome design.
The Experiment: Prompting a Genome Model
The team fine-tuned Evo 1 and Evo 2 on over 2 million bases of bacteriophage DNA, then on Microviridae sequences. They prompted the models with 4–9 bases from the start of the ΦX174 genome. The models output 302 candidate genomes, of which 285 were synthesized and tested in E. coli. Sixteen inhibited bacterial growth, confirming they were functional viruses.
Viability: 5.6% Success Rate, but High Similarity Helps
Only 16 of 285 (5.6%) outputs were viable. However, among sequences with ≥98% similarity to the original ΦX174, viability jumped to 46%. This aligns with known fragility: a single amino acid change has a 20% chance of inactivating the virus. Yet the AI produced viable viruses with far more changes than expected—nearly a quarter of those with >25 amino acid changes were viable, including two with over 50 changes. Random mutation would predict near-zero viability in that range.
Distinct Features: Gene Loss, New Genes, and Swaps
Some AI-designed viruses lost an entire protein, compensated by other changes. One added a new gene. Others had genes of different lengths. One swapped a gene with a distantly related virus. These features would be challenging to achieve through natural evolution.
Phage Therapy Cocktail: AI Viruses Overcome Resistance
The team tested a cocktail of the 16 viable AI viruses against E. coli strains resistant to natural phages. While a natural phage cocktail failed, the AI-generated cocktail quickly evolved to infect the resistant hosts. The researchers suspect DNA segment swapping and additional mutations played a role.
Risks and Governance
The researchers excluded vertebrate viruses from training data, but they warn that someone with sufficient compute could include them. They call for better governance of AI and custom DNA ordering. The paper ends with a call to action.
What This Means for Developers
This is a proof-of-concept that large genome models can generate functional biological sequences. While the immediate application is phage therapy, the same techniques could apply to other biomolecules. For developers, this highlights the power of transformer-based models beyond text—they can learn the 'language' of DNA.
Next Steps
If you're working on AI for biology, consider how these models can be fine-tuned for specific tasks. The code and data from this study are likely to be released; watch for them. And if you're ordering synthetic DNA, be aware of the dual-use concerns.
The Fine Print
The study used Evo 1 and Evo 2, which were trained on a diverse set of genomes. The models were originally trained to predict the next base in DNA sequences, similar to language models predicting the next word. The team's pre-processing included discarding sequences with long homopolymers or unusual GC content.
Editor's Take
I've been following the Evo series since the first paper, and this is the first time I've seen a large genome model produce something that actually works in the lab. The 5.6% success rate is low, but the fact that the AI can generate viable viruses with dozens of changes is remarkable. I'm curious to see if the approach scales to more complex genomes. The governance call is timely—we need to think about biosecurity before this becomes mainstream.
Developer Insights
- Fine-tuning matters: The models were fine-tuned on specific phage families, dramatically improving output relevance.
- Constraints are key: Pre-filtering by sequence length, GC content, and spike protein similarity improved hit rate.
- Test in the lab: Computer predictions alone aren't enough—synthesis and biological testing are essential.
Quiz
(No quiz provided in this article.)
