AI Can Now Write Entire Genomes — And Some of Them Actually Work
Stanford researchers used genome language models to design complete bacteriophages. Sixteen of those AI-generated genomes produced viable viruses in the lab.
For years, generative AI has mostly been associated with things that live inside computers.
Text. Images. Music. Code.
Now the same basic idea is starting to move into something much more physical: biology itself.
Researchers from Stanford University and the Arc Institute have demonstrated that genome language models can generate complete DNA sequences for bacteriophages — viruses that infect bacteria — and that some of those designs actually work when synthesized in a laboratory.
This is not merely AI predicting whether a mutation might be useful.
The model generated entire genomes.
Researchers synthesized selected designs.
And some of them became functioning biological entities capable of infecting and killing bacteria.
That changes the conversation around generative AI quite dramatically.
From Language Models to Genome Models
Large Language Models work by learning statistical structure from sequences of tokens.
Give them enough examples of language and they become surprisingly good at predicting what should come next.
DNA can also be represented as a sequence.
Instead of words and punctuation, the alphabet is mostly:
A, C, G, and T.
That makes genome modeling look conceptually familiar to anyone who has worked with modern language models.
The implementation, of course, is considerably more complicated.
One of the models behind this research is Evo 2, a biological foundation model designed to learn patterns across genomic sequences.
The larger Evo 2 model contains around 40 billion parameters and was trained on roughly 9.3 trillion DNA tokens, using genomic data spanning bacteria, archaea, eukaryotes, and bacteriophages.
Its context window can reach one million nucleotides.
In other words, instead of learning grammar from human language, Evo 2 learns something closer to the statistical grammar of DNA.
And this distinction matters.
Evo 2 is not ChatGPT with DNA pasted into the prompt.
It is a foundation model specifically built around genomic sequences.
The Experiment: Can AI Write a Working Genome?
The researchers focused on a bacteriophage called ΦX174.
Bacteriophages are viruses that infect bacteria rather than humans. ΦX174 is particularly useful for research because its genome is tiny by biological standards — roughly 5,400 base pairs.
Compare that with the human genome, which contains more than 3 billion.
The researchers used genome language models to explore new phage genomes based around ΦX174.
The models produced a huge number of potential sequences.
From those candidates, researchers selected 285 designs for physical synthesis and laboratory testing.
Then came the important part.
Did any of these computer-generated DNA sequences actually produce functioning viruses?
Yes.
Sixteen did.
Those phages were capable of infecting E. coli, replicating, and producing measurable biological effects.
Some also showed fitness characteristics that compared favorably with the natural ΦX174 bacteriophage.
This is the part worth paying attention to.
The model did not simply generate DNA that looked realistic.
Some of its outputs crossed the boundary from plausible sequence to functional biological system.
Why This Is More Interesting Than “AI Made a Virus”
The obvious headline is:
AI generated new viruses.
Technically true.
But it misses the larger point.
The interesting breakthrough is that generative models are beginning to learn biological constraints well enough to design functioning systems at the genome scale.
Until now, a lot of AI-driven biology has focused on narrower problems:
- predicting protein structure,
- estimating the effects of mutations,
- designing individual proteins,
- identifying drug candidates,
- analyzing genomic variants.
Those are important achievements.
But a genome is not a single component.
Genes interact with regulatory regions, structural elements, proteins, replication machinery, and one another.
Changing one piece can break something somewhere else.
That makes whole-genome generation a much harder problem than generating an isolated biological sequence.
The new research suggests genome models are beginning to capture enough of those relationships to produce viable designs.
The Science paper describes the work as the first generative design of complete bacteriophage genomes using genome language models.
That is a meaningful milestone.
The Antibiotic Angle May Be the Most Important One
There is also a very practical reason researchers care about bacteriophages.
Antibiotic resistance is getting worse.
Bacteria evolve.
A drug that kills a bacterial population today can gradually become less effective as resistant strains survive and spread.
Phage therapy offers a different approach.
Instead of using a chemical antibiotic, doctors could potentially use viruses that specifically infect harmful bacteria.
The problem is familiar: bacteria can evolve resistance to phages too.
This is where AI-generated diversity becomes interesting.
Rather than relying on one bacteriophage, researchers can potentially create genetically diverse phage cocktails that attack bacteria in different ways.
In the Stanford experiment, a mixture of generated phages was able to overcome E. coli populations that had developed resistance to natural ΦX174.
That does not mean AI-generated phage treatments are ready for hospitals.
They are not.
But it demonstrates a potentially powerful strategy: use generative models to explore biological design spaces much faster than humans could manually.
Evo 2 Is Basically Autocomplete for DNA — But That Analogy Has Limits
It is tempting to describe a genome language model as:
ChatGPT for DNA.
As a mental model, that works surprisingly well.
Both systems learn statistical relationships in long sequences.
Both can predict plausible continuations.
Both can generate new sequences that resemble structures found in their training distributions.
But biology has one enormous difference.
When an LLM produces a bad paragraph, you delete it.
When a biological model produces a genome, that sequence can potentially be synthesized and tested in the physical world.
Get Vitaly Sem’s stories in your inbox
Join Medium for free to get updates from this writer.
The output is no longer just information.
It can become matter.
That means evaluation changes too.
A text model can be judged using benchmarks, human feedback, or task accuracy.
A genome model eventually needs experimental validation.
The laboratory becomes part of the inference pipeline.
That is a very different kind of AI system.
No, This Does Not Mean AI Can Design Humans Now
The leap from a tiny bacteriophage to something like the human genome is enormous.
ΦX174 contains only a few thousand DNA bases.
The human genome contains billions.
The complexity difference is not merely linear.
Human biology involves regulatory systems, development, epigenetics, cell specialization, environmental interaction, and countless mechanisms we still do not fully understand.
Evo 2 itself can process much longer sequences than previous genome models, and researchers are already exploring genome-scale generation across different organisms.
But “AI generated a working bacteriophage genome” and “AI can design a human genome” are nowhere near the same statement.
Even the researchers behind the project describe controllability and greater genetic novelty as major open research problems.
So no — we are not one software update away from downloading custom humans from GitHub.
Biology remains annoyingly resistant to product roadmaps.
And Then There Is Biosecurity
This technology obviously creates uncomfortable questions.
If AI becomes better at designing biological systems, what prevents the same technology from being used for harmful purposes?
The Evo 2 team anticipated this issue.
Its training dataset deliberately excluded genomic sequences from viruses that infect eukaryotic hosts, including viruses affecting animals and humans.
The researchers also found that this exclusion significantly reduced the model’s performance on those viral sequences.
That is an important safeguard.
It is not a permanent solution.
The larger problem is that biological AI is improving faster than many existing oversight mechanisms were designed to handle.
Traditional biosecurity often focuses on known pathogens, known DNA sequences, known dangerous organisms, and controlled laboratory procedures.
Generative biology introduces something new: sequences that may never have existed before.
That makes screening fundamentally harder.
A future biosecurity system cannot simply ask:
“Is this sequence on the dangerous list?”
It may also need to ask:
“What could this sequence do?”
That is a far more difficult computational problem.
Researchers commenting on the work have already argued that biosecurity frameworks need to evolve alongside AI genome design rather than after it.
Open Source Makes the Question Even More Interesting
Evo 2 is not a closed internal research system.
Its model weights, inference code, training code, and other resources have been released openly.
From a scientific perspective, that makes sense.
Open models allow researchers to reproduce findings, inspect behavior, build better tools, and accelerate biomedical research.
The same openness also creates an unavoidable tension.
The more capable biological foundation models become, the harder the discussion around unrestricted access becomes.
We have already seen this debate with general-purpose AI.
Biology raises the stakes because model outputs can eventually leave the computer.
The argument is no longer purely about generated misinformation, malware, or automation.
It is about systems capable of proposing changes to living organisms.
There probably will not be a simple answer.
Completely restricting the technology could slow potentially life-saving research.
Completely ignoring the risk would be equally naive.
The likely future is a combination of model-level safeguards, DNA synthesis screening, laboratory oversight, access controls for certain capabilities, and better computational methods for predicting biological risk.
Biology Is Becoming Programmable
The most important thing about this research is not the 16 bacteriophages.
It is the direction of travel.
Computing has traditionally helped us observe biology.
First we sequenced genomes.
Then we built tools to analyze them.
Machine learning allowed us to predict their properties.
Now generative models are beginning to design them.
That is a fundamental shift.
We are slowly moving from: read biology
to: understand biology
to: write biology.
Software engineers are familiar with what happens when a system becomes programmable.
Iteration accelerates.
Search spaces become easier to explore.
Automation appears.
Abstractions improve.
Eventually, things that once required specialists performing painstaking manual work become accessible through higher-level tools.
Biology may be entering the early stages of the same transition.
Except the runtime is a living cell.
And production bugs become considerably more interesting.
Final Thoughts
Generative AI spent the last few years learning to write prose, generate images, compose music, and produce software.
Writing functional genomes is different.
For the first time, genome language models have demonstrated that AI-generated complete viral genomes can be converted from digital sequences into functioning biological systems.
The immediate application — better bacteriophages against antibiotic-resistant bacteria — could be extremely valuable.
The broader implication is much larger.
AI is beginning to move from describing living systems to designing them.
That does not mean synthetic organisms on demand are around the corner.
It does mean the boundary between computer science and biology is becoming increasingly difficult to see.
And once biology becomes another programmable medium, AI stops being purely a software story.
It becomes a life sciences story too.
References
Samuel H. King et al. -Generative design of bacteriophages with genome language models, Science, 2026.
Garyk Brixi et al. -Genome modelling and design across all domains of life with Evo 2, Nature, 2026.
Stanford University -AI designs a novel E. coli killer, 2026.
