⏱️ Lectura: 11 min
An artificial intelligence model wrote, letter by letter, the complete DNA of a virus, and that virus worked. Arc Institute announced that its genomic model Evo generated the genomes of dozens of candidate bacteriophages, and that after synthesizing and testing them in the lab, 16 turned out to be viable: they infected real bacteria and replicated inside them.
📑 En este artículo
The news, reported by Wired, isn’t about a human virus or a biological weapon: these are bacteriophages, viruses that only attack bacteria. It’s one of the first cases of AI-designed viruses that work outside of theory, and it marks a shift in category: moving from AI that predicts existing biological structures to AI that designs entire organisms from scratch.
TL;DR
- Arc Institute used AI to generate the complete genome of AI-designed viruses: 16 bacteriophages turned out to be functional, according to Wired.
- The model used, Evo, treats DNA as language: it predicts the next base (A, T, C, or G) the same way an LLM predicts the next word.
- The generated viruses are bacteriophages that infect E. coli, not human pathogens: a test system that has historically been safe in biology.
- phiX174, the first genome ever sequenced (1977), is the same type of small bacteriophage used as a reference in these experiments.
- The full process: generate the sequence with AI, synthesize the DNA on order, insert it into bacteria, and confirm that the phage infects and replicates.
- The case reopens the biosecurity debate: DNA synthesis providers screen orders to block sequences from known pathogens.
- Evo is an open source project: its code and weights are publicly available for research on GitHub and Hugging Face.
What Happened
The experiment is built on Evo, a language model trained not on text but on DNA sequences from bacteria, archaea, and viruses. Instead of predicting the next word in a sentence, Evo predicts the next nitrogenous base (A, T, C, or G) in a genetic chain. That mechanism lets it generate, base by base, a complete and coherent DNA sequence, the same way a text model generates an entire paragraph.
The Arc Institute team used Evo to generate the complete genome of bacteriophages, viruses with small genomes (thousands of bases, not millions) that infect bacteria like Escherichia coli. From a set of sequences generated by the model, the researchers synthesized the most promising ones as physical DNA, inserted them into bacteria, and observed which ones managed to produce viral particles capable of infecting and killing their host. Sixteen of those AI-designed genomes turned out to be functional. Arc Institute describes the result as one of the first demonstrations of AI-designed viruses behaving like real viruses in the lab, not just plausible sequences on a screen.
💭 Key point: designing a protein (what AlphaFold does) is not the same as designing an entire organism. A functional viral genome needs dozens of genes, promoters, and regulatory signals to fit together simultaneously for the virus to assemble and replicate.
Context and History
Bacteriophages have been at the center of experimental biology for more than a century because they’re simple, safe to handle, and quick to test: a bacterial culture on a lab plate shows within hours whether a phage infects or not. That’s why phiX174, a bacteriophage with a genome of just a few thousand bases, was the first complete genome ever sequenced, in 1977, by Frederick Sanger’s team. Half a century later, that same small and well-understood type of virus serves as the testing ground for the first generation of AI-designed viruses.
Evo is the evolution of a model lineage that started in 2024 with Evo 1, trained on genomic sequences from prokaryotes (bacteria and archaea) and some viruses. Evo 2 expanded training to eukaryotic genomes, including fragments of the human genome, to capture evolutionary patterns shared across very different species. The logic is the same one that made text LLMs popular: the more sequences and the more diverse they are during training, the better the model generalizes the patterns that make a DNA sequence biologically valid instead of just random.
The immediate precedent for this kind of work is AlphaFold, DeepMind’s system that predicts a protein’s three-dimensional structure from its sequence. AlphaFold reads biology: it turns a sequence into a shape. Evo, on the other hand, writes biology: it turns a design intent into a new sequence that didn’t exist before. That leap, from predicting to generating, is what makes these models a tool as interesting as it is delicate.
Technical Details and Performance
A genomic model like Evo runs on the same architecture that underpins text LLMs: a network that processes long sequences and learns to predict the next token from context. The difference is the vocabulary: instead of tens of thousands of words and subwords, Evo’s vocabulary has four symbols, A, T, C, and G, plus a few special control tokens. Generating a complete genome means making that prediction, base by base, tens of thousands of times in a row without losing the sequence’s overall coherence.
The experiment’s workflow has four steps: computational generation of candidates, physical synthesis of the DNA, insertion into the host bacterium, and verification of infectivity on a plate. Here’s a summary:
flowchart TD
A["Evo generates candidate sequences"] --> B["Computational filtering by similarity to known phages"]
B --> C["Physical DNA synthesis on order"]
C --> D["Insertion of the genome into E. coli"]
D --> E{"Do viral particles form?"}
E -- "Yes" --> F["Functional phage: 16 confirmed cases"]
E -- "No" --> G["Candidate discarded"]In practice, this is what it conceptually looks like when using a genomic model to generate a sequence (simplified example, not production-ready):
from evo2 import Evo2
modelo = Evo2("evo2_7b")
prompt_adn = "ATGAAACGCATTAGCACCACC" # known initial fragment
secuencia_generada, _ = modelo.generate(
prompt_seqs=[prompt_adn],
n_tokens=5000,
temperature=0.7,
)
print(secuencia_generada[0])Before synthesizing any sequence generated by a model like this, the mandatory step is comparing it against databases of known pathogens. A real example of that verification using standard bioinformatics tools:
blastn -query candidato_fago.fasta \
-db nt \
-remote \
-outfmt "6 qseqid sseqid pident length evalue" \
-max_target_seqs 5That command compares the generated sequence against the public GenBank database and returns how similar it is to already-cataloged organisms. It is, in essence, the same verification logic used today by commercial DNA synthesis providers before fulfilling an order: if the sequence looks too similar to a pathogen on the control list, the order gets blocked.
How to Try It
Evo is an open source project. The repository and model weights are published by Arc Institute so other research labs can reproduce the approach, train their own variants, or use it as a base for narrower tasks, like designing variants of an existing protein instead of an entire genome. Running the full version of Evo 2 requires a GPU with considerable memory, but smaller versions of the model exist, with fewer parameters, designed for experimenting without needing a cluster.
- Repository: the training and inference code is published on Arc Institute’s GitHub.
- Model weights: available for download on Hugging Face under a research license.
- Hardware requirement: the large checkpoints need a GPU with tens of GB of VRAM; the small checkpoints run on a single consumer GPU.
💡 Tip: before running local inference, check Arc Institute’s official repository: they usually publish example notebooks ready to run in a cloud GPU environment, without installing anything locally.
Impact and Analysis
The result matters for two opposing reasons regarding the future of AI-designed viruses. The first is therapeutic: bacteriophages are a real alternative to antibiotics against resistant bacteria, and being able to design custom phages to target a specific bacterial strain accelerates a field, phage therapy, that today depends largely on finding the right phage in nature through trial and error. The second reason is biosecurity: if a model can generate the complete genome of a functional virus based on learned patterns, the inevitable question is how far that same capability is from being applied to organisms dangerous to humans.
| Biological design approach | When it’s used | Advantage | Limitation |
|---|---|---|---|
| Classical genetic engineering | Modifying a specific gene in an existing organism | Predictable, decades of practice | Doesn’t generate new organisms from scratch |
| Directed synthetic biology | Assembling genetic circuits with known parts | Fine control over each component | Requires expert manual design, slow |
| Generative models like Evo | Exploring variants or complete genomes at scale | Generates thousands of candidates in hours | Most candidates aren’t viable; requires lab validation |
The organizations that manage pathogen registries and DNA synthesis providers have spent years working with lists of controlled sequences to block orders that match known dangerous agents. The problem with a generative model is that it can produce a sequence that’s functionally similar to a pathogen without being textually identical to any entry on that list, which complicates automatic filtering based on sequence matching.
⚠️ Note: the bacteriophages used in this experiment don’t infect humans and have been microbiology’s safest test organism for more than a hundred years. The biosecurity debate is about the trajectory of the capability, not an immediate risk from this specific experiment.
What’s Next
The logical next step for this type of research is scaling from phages with a few thousand bases to larger and more complex viral genomes, always within organisms that aren’t pathogenic to humans, and refining the biosecurity filters that run before synthesis. In parallel, regulators and synthetic DNA providers are expected to review their screening protocols to incorporate detection based on functional similarity rather than just exact sequence matching, precisely because of the kind of case this experiment puts on the table.
For phage therapy, the immediate path is testing whether AI-designed phages can target antibiotic-resistant bacterial strains with more precision than phages isolated from natural environments like wastewater, which is how most therapeutic phages are discovered today.
A real limitation of this work: the 16 phages confirmed as functional have genomes of just thousands of bases, orders of magnitude simpler than a virus capable of infecting humans, which also needs machinery to evade a complex immune system. Going from designing a phage that infects E. coli on a plate to designing a viable human pathogen isn’t an incremental step, it’s a massive leap in biological complexity. That doesn’t eliminate the underlying concern, but it does temper the sensationalist reading of the headline.
📖 Summary on Telegram: View summary
Try it yourself: go to Arc Institute’s repository on GitHub and check out the example notebooks to generate your first DNA sequence with an open genomic model.
Frequently Asked Questions
What is a bacteriophage?
It’s a virus that exclusively infects bacteria, not human or animal cells. They’ve been used for more than a century as a study system in microbiology because they’re simple and safe to handle in the lab.
Did Evo also design viruses that can infect humans?
No. The experiment reported by Wired was limited to bacteriophages that infect E. coli, a standard model organism in biology, unrelated to human pathogens.
How is Evo different from AlphaFold?
AlphaFold predicts a protein’s three-dimensional shape from its sequence. Evo generates complete, new DNA sequences, including entire genomes, instead of just predicting structures of existing sequences.
Is it dangerous that this technology exists?
The risk depends on scale: generating a small phage isn’t the same as designing a complete human pathogen, but the biosecurity community is calling for stronger screening of DNA synthesis orders against generative models.
Can I use Evo myself?
Yes, the model is open source and its code is published by Arc Institute, although running the full version requires a GPU with considerable memory.
What practical application does designing phages with AI have?
Phage therapy: using specific viruses to eliminate antibiotic-resistant bacteria, an alternative that’s increasingly studied in response to rising bacterial resistance.
References
- Wired: the original coverage of the experiment with Evo and the 16 functional bacteriophages.
- Arc Institute: the research organization that developed the Evo genomic model.
- Wikipedia: Bacteriophage: general context on what bacteriophages are and how they’re used in research.
- Wikipedia: PhiX174: the history of the first complete genome ever sequenced, in 1977.
- Arc Institute’s GitHub: the open source repository for the Evo project.
📱 Do you like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Logan Voss en Unsplash
0 Comments