AI De Novo Designs 16 Novel Viable Bacteriophages: Not Prediction, but Creation
> King et al., Nature (2026) — *Generative design of novel bacteriophages with genome language models* > Authors: Samuel H. King, Claudia L. Driscoll, David B. Li, Daniel Guo, Aditi T. Merchant, Garyk Brixi, Max E. Wilkinson, Brian L. Hie > Institutions: Arc Institute, Stanford University > DOI: 10.1101/2025.09.12.675911
---
1. How extraordinary is this, really?
Imagine being handed books written in a language you never studied — but where you somehow learn what word plausibly comes next on each page. After reading millions of books, one day you decide to write one yourself. Not copying, not adapting — composing from a blank page, word by word.
Then you hand your book to a native speaker. Not only can they read it — they find passages they could not have written themselves in their native tongue.
That is what Evo did.
Evo is a DNA language model. Unlike AlphaFold, it doesn't predict protein structures; unlike GPT-4, it doesn't write poetry. It learned the grammar of life — not English grammar, not Python syntax, but DNA grammar: how the four letters (A, C, G, T) combine across sequences hundreds of thousands to millions of bases long into something alive and self-replicating.
Then it wrote one from scratch.
Not a rewrite, not a recombination of natural sequences — a de novo genome of 5–6 kilobases with 11 genes, regulatory elements, and plausible coding density. And 16 of these genomes actually came alive in the lab. They formed plaques on bacterial plates, lysing bacteria, replicating, and propagating like real viruses.
The success rate was 5.3% (16 viable out of 302 designs). That sounds low — but in a biological context, it is astonishingly high.
---
2. Evo: the language model of life
2.1 Why DNA language models are harder than other bio foundation models
Protein language models (ESM2, ProGen) have already demonstrated strong predictive power. But proteins are only one dimension of life. DNA is the more fundamental language — it encodes RNA, proteins, regulatory elements, and all the interactions among them.
The ΦX174 phage genome is only 5.4 kb with 11 genes, yet it is a complete living system. This means:
- Every gene's open reading frame (ORF) must start and stop correctly
- Regulatory elements must sit at the right relative positions
- Genes must not conflict through overlaps
- Coding density must be high (ΦX174 exceeds 95%)
- All proteins must physically interact to form functional particles
- Local grammar: patterns of transcription factor binding sites, promoters, ribosome binding sites
- Mid-range structure: intron/exon boundaries (in eukaryotes), operon architecture
- Long-range dependencies: how distal enhancers affect gene expression tens of thousands of bases away
- Cross-modal associations: how DNA sequences encode proteins, and how proteins interact with RNA/DNA
- Fitness effects of E. coli protein mutations (competitive with dedicated protein LMs)
- Fitness effects of non-coding RNA mutations (surpassing dedicated RNA LMs)
- Whether promoter–RBS combinations drive gene expression
- Essential genes in bacteria or phages — with no supervision at all
- ΦX174's J protein: 38 amino acids
- G4's J protein: 25 amino acids
- G4's genome is only 63% similar to ΦX174
- Prior experimental work showed that grafting G4's J protein into ΦX174 yields a non-viable phage
- Transcription factor binding sites: a few base pairs
- Genes: hundreds to thousands of base pairs
- Operons: thousands to tens of thousands
- Genomic islands / prophages: tens to hundreds of thousands
- All three domains of life: bacteria, archaea, eukaryotes
- Viruses (non-eukaryotic only, for safety reasons)
- Plasmids and other mobile genetic elements
- Growth competition: several AI-designed phages outcompeted natural template ΦX174
- Lysis kinetics: faster bacterial OD600 decline after infection, indicating higher lysis efficiency
- Cryo-EM: the complete Evo-Φ36 virion was resolved at 2.9 Å, proving AI-designed proteins assemble correctly
- Resistance breaking: an AI-designed phage cocktail fully suppressed growth of ΦX174-resistant strains within five passages
- Gene synthesis (~$100–500 per construct)
- Cloning, transformation, culturing
- Plaque assays and confirmation
- Cell-free systems: rapid screening of genome activity in vitro, without living cells
- ML-assisted experimental design: Bayesian optimization or active learning to select the most informative candidates
- Higher-fidelity simulation: end-to-end modeling from gene sequence to proteome to cellular dynamics
- We cannot predict which designs will succeed using traditional structural-biology intuition
- Every success is like a small scientific discovery — requiring experimental validation
- Confidence calibration of computational models remains an open problem
- Training data filtering: excluding all eukaryote-infecting viruses (including human pathogens)
- Experimental host: non-pathogenic E. coli C, not a clinical pathogen
- Biosafety level: BSL-2 (standard biosafety cabinet work)
- Sequence novelty: designed sequences differ substantially from natural phages, reducing recombination risk with known pathogens
- Synthetic biology community norms: following iGEM and DIY Bio safety guidelines
- What if the model is maliciously fine-tuned on pathogenic virus data?
- What if designed phages recombine with wild phages in the environment?
- What if AI designs sequences with broader host ranges?
- Phages are nature's bacterial predators, evolved over billions of years
- They are highly specific — infecting only particular bacterial strains, sparing human commensal flora
- But natural phage therapy faces narrow spectrum and resistance evolution
- AI-designed phage cocktails overcome resistance: in experiments, bacteria failed to evolve resistance within five passages
- Programmable host range: host specificity can be tuned by controlling spike protein sequence similarity
- Design space beyond natural diversity: AI can explore sequence space evolution never visited
- How big is the risk?
- What experiments can quantify it?
- What is the benefit?
- 1952: The Hershey–Chase experiment proves DNA is the genetic material
- 2003: Human Genome Project completed — we first "read out" the book of life
- 2012: CRISPR-Cas9 — we learned to "edit" life
- 2026: Evo — for the first time we "wrote" an entirely new living system
- King et al. (2025/2026). *Generative design of novel bacteriophages with genome language models*. bioRxiv 2025.09.12.675911; Nature.
- Nguyen et al. (2024). *Sequence modeling and design from molecular to genome scale with Evo*. Science.
- Arc Institute. (2024). *Evo: DNA foundation modeling from molecular to genome scale*. https://arcinstitute.org/news/evo
- NVIDIA Research. (2025). *The Illustrated Evo 2*. https://research.nvidia.com/labs/dbr/blog/illustrated-evo2/
- Poli et al. (2023). *StripedHyena: Efficient Transformer Substitutes*. Together AI.
- Hazy Research, Stanford. (2024). *Learning from DNA: a grand challenge in biology*. https://hazyresearch.stanford.edu/blog/2024-03-14-evo
This isn't predicting a single protein fold. It's predicting whether a tiny operating system can boot.
2.2 What Evo learned
Evo was trained on 300 billion nucleotides from 2.7 million prokaryotic and phage genomes. It learned:
Most striking are the zero-shot capabilities — without fine-tuning on any labeled data, Evo can predict:
This suggests Evo learned not statistical correlations but the causal structure of biology.
2.3 Key experiment: Evo-Φ36's J protein
This is the finding that gave the original author goosebumps.
Evo-Φ36 is an AI-designed phage whose genome is 96.9% identical in nucleotides to template ΦX174 — apparently very conservative. But look closer: its J protein was replaced with the J protein from another phage, G4.
The J protein is a DNA packaging protein that also props up the capsid.
Yet Evo-Φ36 was viable.
Cryo-EM structure determination (2.9 Å resolution, comparable to ΦX174's 2.8 Å) shows that although Evo-Φ36's J protein is 13 amino acids shorter and lacks a key capsid-binding domain, it found a novel protein–protein co-evolution solution in coordination with the rest of the genome's context. Evo didn't just swap a part — it redesigned the entire assembly logic.
This isn't "assembling LEGO." It's understanding a complex system's dynamics and then redesigning it.
---
3. The StripedHyena 2 architecture: three key numbers
Evo's core is the StripedHyena architecture — a hybrid design built for ultra-long sequences. Its philosophy: let different layers do different things.
3.1 Number one: 131,072 — context length
Evo 1's context window is 131,072 tokens — and each token is one nucleotide. Evo can process 130k bases of continuous DNA in one shot — enough to cover the entire ΦX174 genome (5.4 kb) with plenty of headroom, plus larger phage and bacterial genome segments.
Why does this matter? Biological features span enormous scales:
A context of only a few thousand tokens would see only local patterns. The 131k context lets Evo learn both local features and long-range regulation simultaneously.
3.2 Number two: 9.3 trillion — Evo 2's training tokens
Evo 2 (the 40B version) was trained on 9.3 trillion nucleotides — 31× Evo 1 (300B tokens).
Training data covered:
What's wild about this number: it far exceeds conventional "Chinchilla-optimal" compute. The Evo team found that StripedHyena keeps improving well beyond the compute-optimal training budget — unlike Transformers, which tend to overfit or show diminishing returns.
3.3 Number three: three Hyena operators + attention across 32 layers
Evo 2-7B's 32 layers are not a homogeneous stack. It uses a multi-hybrid block layout, where each layer contains one of four operators:
| Operator | Role | Biological analogue | |------|------|----------| | Hyena SE (Short Explicit) | Short convolutional filters for local features | TF binding sites, restriction sites | | Hyena MR (Medium Regularized) | Medium-length convolutions | Exon/intron boundaries, intergenic regions | | Hyena LI (Long Implicit) | Implicit long convolutions | Distal enhancers, genomic islands | | MHA (Multi-Head Attention) | Global information fusion | Cross-gene regulation, system-level coordination |
Key insight: this is not a "faster Transformer" engineering exercise, but a bio-inspired design with layer specialization — like the visual cortex's V1 (edges), V2 (textures), V4 (shapes), IT (objects), different StripedHyena layers specialize in DNA patterns at different scales.
MHA layers make up only about 10% (3/29 in Evo 1, similar ratios in Evo 2), but they are the hubs of information fusion — integrating local and long-range features into a unified genome representation.
This design lets Evo 2-7B process 131k sequences more than 3× faster than a Transformer in training, while maintaining better perplexity — which directly determines how deeply the model understands DNA sequences.
---
4. A three-act experiment: AI brainstorms → computational screening → reality check
Act I: AI brainstorms (generation)
Input: The 5' region of ΦX174 as prompt (lengths from 10 bp to 500 bp) Model: Evo 1 SFT (supervised fine-tuned) and Evo 2 SFT Operation: Autoregressive generation with temperature sweeps (0.7–1.2) Output: 1,000 sequences per configuration, thousands of candidate genomes total
The key variable is generation temperature. Low temperature (0.7) makes the model conservative, producing sequences closer to natural phages in training data; high temperature (1.2) makes it bold and divergent — but likelier to produce biological nonsense.
The team found that moderate temperature (~0.9) and moderate prompt length (~100 bp) hit the sweet spot: preserving genomic-architecture plausibility (architecture similarity score > 0.38) while achieving high sequence novelty.
Act II: computational screening
From the thousands of sequences, the team built a six-layer filtering framework:
1. Viral classification (geNomad): confirm sequences are classified as viral, not bacterial or noise 2. Novelty detection (nucleotide BLAST): low query coverage and similarity to training data (Microviridae) 3. Quality assessment (CheckV): >87% rated "High Quality" or "Complete" 4. Gene annotation (PHROGs database): predicted protein functions match known phage genes 5. Structural plausibility (ESMFold): predicted structures' pLDDT comparable to natural proteins 6. Architecture similarity (custom algorithm): compare to ΦX174's genome architecture to ensure sensible ORF layout
These six layers narrowed thousands of sequences down to 302 — each passing the computational test of "this looks like a plausible phage genome."
Act III: reality check (experimental validation)
All 302 sequences were synthesized, cloned into plasmids, electroporated into E. coli C, and assayed for plaques.
Result: 16 formed plaques (5.3%).
But 5.3% was just the start. The 16 survivors faced tougher tests:
---
5. Three technical bottlenecks behind the 5.3% success rate
A 5.3% success rate sounds unimpressive — but it is not Evo's failure; it is a mirror of biology's complexity.
Bottleneck 1: genomes are coordinated systems, not parts lists
A phage genome is not a linear sum of 11 independent genes. Expression levels, translation efficiency, protein folding, protein interactions — all are coupled.
Evo-Φ36 illustrates this best: G4's J protein alone is lethal in ΦX174, but in Evo's redesigned context, other genes' regulation and expression compensated for the difference. Evo learned co-design — but that means one gene's "error" can be compensated by another's "adjustment," and such compensatory relationships cannot be fully predicted by computational screening.
Essentially: this is an NP-hard combinatorial optimization problem. 302 candidates cover a tiny corner of the design space.
Bottleneck 2: the hard ceiling of experimental throughput
302 designs is already the limit of high-throughput synthesis. Each design requires:
At a 5% success rate, finding one survivor requires testing ~20 on average — an expensive game of luck. The team tested 302 to find 16, showing that computational filtering works but still produces many false positives.
Future breakthrough points:
Bottleneck 3: context-dependent structural compatibility
Cryo-EM data show that Evo-Φ36's J protein's N-terminal 14 amino acids are unresolvable inside the capsid — indicating this segment is flexible/disordered, quite unlike ΦX174's rigid binding mode.
This means Evo didn't design "the standard answer" — it found an alternative solution outside known constraints. That creativity is stunning, but it also means:
---
6. Biosafety vs. the post-antibiotic era: which side are you on?
6.1 Safety guardrails
The Evo team was clear-eyed about safety. Their safeguards include:
But the guardrails are not perfect:
6.2 A beacon for the post-antibiotic era
On the other hand, phage therapy is one of the most promising answers to antibiotic resistance:
This work's breakthrough:
6.3 So which side?
In the Feynman style: I don't take sides. I look at the experimental results.
Safety is a real concern. But "don't do it because it might be risky" is not a scientific attitude — it's an attitude of fear. The right questions are:
Antibiotic resistance kills roughly 700,000 people a year globally. By 2050, that is projected to reach 10 million.
If AI-designed phages can turn even one drug-resistant infection from "untreatable" into "treatable with a phage" — that benefit deserves serious consideration.
But the precondition: safety validation must be the first priority. Not an afterthought, not a one-line mention in a paper — but demonstrated with experimental data, like this work's cryo-EM.
---
7. What order of breakthrough is this?
Its place in the history of biology:
Note the keywords: entirely new and alive.
Previously we engineered viruses — inserting GFP into lentivirus, putting Cas9 into AAV. That was recombination, rearranging parts nature already made.
This time it's de novo design. Some genes in Evo's genomes have no significant homologs in natural databases. It isn't imitating nature — it is exploring corners nature never visited.
This is not Cargo Cult Biology — mimicking form without substance. This is substance.
---
8. Unanswered questions (honestly)
1. How much "real" biology did Evo learn, versus statistical pattern matching? Evo-Φ36's success looks like an understanding of structural constraints, but we have no mechanistic explanation — we don't know whether it "knows" the J protein must interact with the capsid in some way, or just got lucky with the sequence.
2. How high can the success rate go? 5.3% is remarkable for basic research but far from sufficient for clinical applications. A 100×–1000× improvement is needed.
3. What about larger genomes? ΦX174 is only 5.4 kb. Real therapeutic phages may be 50–200 kb. Evo can handle a 131k context — but designing a functional 200 kb genome is a categorically different challenge.
4. What about eukaryotes? For safety reasons, Evo's training excluded eukaryotic viruses. But phage therapy could in principle work against eukaryotic pathogens (like fungi). Should this path be pursued? How, safely?
---