An AI Wrote a Virus From Scratch, and It Worked

Some of the AI-designed phages beat nature's own version at killing E. coli. That's exciting, and it's also the whole problem.

Welcome Back,

Stanford researchers have synthesised nearly 300 phages from DNA sequences produced entirely by Evo 2, a generative AI model, then tested them in the lab against E. coli, according to AI News. Sixteen of those AI-designed phages showed particularly strong bacteria-killing activity, and some actually outperformed the natural virus they were modeled on.

The work centers on bacteriophage ΦX174, pronounced "FYE-ex-1-7-4," a virus that infects bacteria rather than humans. Brian Hie, the Stanford assistant professor who created Evo 2, and bioengineering graduate student Samuel King led the research, asking the model to do something genuinely difficult, generate an entire viral genome from end to end, in a single pass, rather than just editing an existing one.

Today we look at how Evo 2 actually generated a working genome, why some AI-designed phages beat the natural version, why bacteria have a much harder time developing resistance to a cocktail than to a single phage, and the genuine safety trade-off Hie himself acknowledges in making this tool open-source.

🧬 Writing an Entire Genome in One Pass

Evo 2 generates new DNA sequences starting from a small snippet of a phage genome. For this project, the researchers pushed the model further than a typical edit or extension, they asked it to produce the entire ΦX174 genome end-to-end in a single left-to-right pass.

That "we didn't add anything" detail matters. This wasn't a human editing an existing genome with AI assistance, it was the model generating a full, coherent viral genome on its own, which then had to actually function once synthesised in a lab.

ΦX174 was chosen deliberately as a relatively compact test case. Its genome contains fewer than 6,000 base pairs, compared with roughly 3 billion base pairs in the human genome. Even at that far smaller scale, Hie noted that researchers still face a genuinely difficult task interpreting a 5,400-character DNA sequence gene by gene, underscoring how much of biology remains hard to parse even in a "simple" system.

Evo 2 generated thousands of candidate genomes before researchers narrowed the field for lab testing.

🏆 When the AI Version Beats the Original

The process generated thousands of candidate genomes before the team selected specific sequences for chemical synthesis and laboratory testing. That's the genuinely striking finding here, Hie said some of Evo 2's suggested phages showed higher fitness than the native ΦX174 in actual laboratory testing.

That result reframes what this project is actually testing. It's not just "can a model propose small, local edits to an existing genome," which is a more modest and already-established capability. It's "can a model create entire, viable viral genomes from scratch," a meaningfully bigger claim, and one this specific result offers real evidence for.

🔬 Screening Thousands of Candidates Down to a Testable Few

Generating thousands of candidate genomes is one thing, chemically synthesising and lab-testing every single one is a completely different cost problem. King developed a computational framework specifically to solve that, assessing traits drawn from ΦX174 and related phages before the team committed resources to synthesis.

"One of the main parts of the design framework was figuring out what traits the genomes should have based on ΦX174 and related phages," King said. "The framework involved several key steps: generating genomes using Evo 2, evaluating options based on the design criteria, selecting optimal candidates, synthesising them chemically, and then testing them in the lab to see which genomes worked best." Hie said this approach reduced synthesis costs by concentrating spending on the candidates the team judged most viable, an important practical constraint that defines the actual limits of this method today, AI generation alone isn't sufficient, computational evaluation, chemical synthesis, and lab testing all remain required steps.

Computational screening narrowed the field before any candidate reached the expensive step of chemical synthesis.

🦠 Why a Cocktail Beats a Single Treatment

The researchers deliberately selected more than one E. coli-targeting phage, because bacteria can evolve resistance to a single treatment relatively easily. A mixture of genetically distinct phages is a much harder target to evade all at once.

The results backed that logic up directly, Stanford reports that a cocktail containing the 16 selected phages rapidly overcame resistance in E. coli that was already immune to the native ΦX174. In other words, bacteria that had already learned to defeat the natural virus still struggled against the AI-designed mixture.

Hie said similar work could eventually target methicillin-resistant Staphylococcus aureus, better known as MRSA, and Pseudomonas aeruginosa, which Stanford describes as a leading cause of medically resistant infections acquired in hospitals. Antibiotic resistance is one of modern medicine's genuinely hardest problems, and a design tool that can generate new resistance-fighting cocktails on demand would be a significant addition to that fight.

⚠️ The Open-Source Trade-off Hie Isn't Hiding From

Hie has released Evo 2 as open-source software, meaning any researcher can download the model and use it to design genomes of their own. That decision has, in Stanford's own account of the project, raised real safety and security discussions.

It's worth being direct about this, that's a genuine, unresolved tension, not a settled debate. Hie's own assessment, that existing pathogens are a bigger practical risk than an AI-generated one, is a judgment call rather than a proven fact, and reasonable biosecurity researchers disagree on how much weight to give either side of that trade-off as these generative biology tools keep improving.

Open-sourcing a genome-design model widens both the research benefit and the potential for misuse at the same time.

🧠 AI Spotlight Analysis

This project is a genuinely useful data point in a much bigger question, can AI models trained on biological sequence data actually understand biology well enough to design functioning organisms, not just describe or classify them. The fact that some AI-generated phages outperformed the natural version they were based on is real evidence the answer is, at least in this narrow case, yes.

King's framing of the project's appeal is worth taking at face value too. "One of the most rewarding parts of this project is the creativity Evo 2 allows. New doors in science are now open because of what we can do with these models." That's a genuine research benefit, not hype, the ability to explore genome-design space far faster than manual engineering allows is a real scientific tool.

That question, controllability, is really the crux of both the promise and the risk here. A model that can reliably generate novel, functional genomes on demand is enormously valuable for medicine, and the exact same capability is what makes the open-source safety debate a genuine one rather than a hypothetical.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.