AI Designs 16 Fully New Bacteriophages from Scratch: 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. Just How Wild Is This?

Imagine being handed books written in a language you never studied — you only know, roughly, what the next word on each page should be. After reading millions of books, one day you decide to write one yourself. Not copying, not adapting — writing from a blank page, word by word.

Then you hand your book to a native speaker. Not only do they understand it — they find passages they themselves could not have written.

That is what Evo did.

Evo is a DNA language model. Unlike AlphaFold (protein structure prediction) or GPT-4 (poetry), it learned the grammar of life — not English, not Python, but DNA: how the four letters A, C, G, T combine across sequences of hundreds of thousands to millions of bases to form something alive and self-replicating.

Then it wrote one from scratch.

Not a rewrite, not a splice of natural sequences — a de novo 5–6 kb genome with 11 genes, regulatory elements, and realistic coding density. And 16 of these genomes actually came alive in the lab: they formed plaques on bacterial lawns, lysing cells, replicating, and reproducing like real viruses.

The success rate was 5.3% (16 survivors out of 302 designs). That sounds low — but in a biological context, it is astonishingly high.

---

2. Evo: The Grammar of Life

2.1 Why DNA Language Models Are Harder Than Other Bio-FMs

Protein language models (ESM2, ProGen) have shown powerful predictive ability, but proteins are only one dimension of life. DNA is the lower-level language — encoding RNA, proteins, regulatory elements, and all their interactions.

The ΦX174 phage genome is only 5.4 kb with 11 genes, yet it is a complete living system. That means:

• Every gene's open reading frame (ORF) must start and stop correctly
• Regulatory elements must sit in the right relative positions
• Genes must not overlap in conflicting ways
• Coding density must be high (ΦX174 exceeds 95%)
• All proteins must physically interact to form functional particles

This is not predicting how one protein folds. It is predicting whether a miniature operating system will boot.

2.2 What Evo Learned

Evo was trained on 300 billion nucleotides from 2.7 million prokaryotic and phage genomes. It learned:

• Local grammar: transcription factor binding sites, promoters, ribosome binding sites
• Mid-range structure: operon organization, gene boundaries
• Long-range dependencies: how distal elements affect gene expression tens of thousands of bases away
• Cross-modal associations: how DNA encodes proteins, and how proteins interact with RNA/DNA

Most striking is the zero-shot capability — without fine-tuning on labeled data, Evo can predict:

• 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 and phages (without any supervision)

This suggests Evo learned not statistical correlation but the causal structure of biology.

2.3 Key Experiment: Evo-Φ36's J Protein

The finding that gave me goosebumps.

Evo-Φ36 is an AI-designed phage whose genome is 96.9% nucleotide-identical to template ΦX174 — very conservative, right? Look closer: its J protein was swapped for the J protein of a different phage, G4.

The J protein is a DNA packaging protein that also props up the capsid.

• ΦX174's J protein: 38 amino acids
• G4's J protein: 25 amino acids
• G4 shares only 63% genome similarity with ΦX174
• Prior experimental work showed that grafting G4's J protein into ΦX174 yields a non-viable phage

But Evo-Φ36 was alive.

Cryo-EM structure determination (2.9 Å resolution, comparable to ΦX174's 2.8 Å) showed that although Evo-Φ36's J protein is 13 amino acids shorter and lacks a key capsid-binding domain, it found a entirely new 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 whole assembly logic.

This is not "assembling LEGO." It is redesigning a complex system after understanding its dynamics.

---

3. The StripedHyena 2 Architecture: Three Key Numbers

Evo's core is the StripedHyena architecture — a hybrid architecture designed for ultra-long sequences. Its philosophy: let different layers do different things.

3.1 Number One: 131,072 — Context Length

Evo 1's context length is 131,072 tokens. Note: each token is one nucleotide. Evo can process 130k bases of continuous DNA in one pass — more than enough to cover the entire ΦX174 genome (5.4 kb) with ample headroom for larger phage and bacterial genome fragments.

Why does this matter? Biological features span enormous scales:

• Transcription factor binding sites: a few bases
• Genes: hundreds to thousands of bases
• Operons: thousands to tens of thousands
• Genomic islands / prophages: tens to hundreds of thousands

A context of only a few thousand tokens would capture only local patterns. The 131k context lets Evo learn 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's 300B tokens.

Training data covered:

• All three domains of life: bacteria, archaea, eukaryota
• Viruses (non-eukaryotic only, for safety reasons)
• Plasmids and other mobile genetic elements

The wild part: this far exceeds conventional "Chinchilla-optimal" compute budgets. The Evo team found the StripedHyena architecture keeps improving well beyond the compute-optimal training frontier — unlike Transformers, which tend to overfit or plateau.

3.3 Number Three: Three Hyena Operators + Attention Across 32 Layers

Evo 2-7B's 32 layers are not homogeneous stacking. It uses a multi-hybrid block layout, with each layer containing 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 choice, but a biology-inspired design where layers specialize. Just as the visual cortex has V1 (edges), V2 (texture), V4 (shape), and IT (objects), StripedHyena's layers specialize in DNA patterns at different scales.

MHA layers make up only ~10% (3/29 in Evo 1, similar ratio in Evo 2), but they are the fusion hubs — integrating local and long-range features into a unified genomic representation.

This design lets Evo 2-7B process 131k sequences over 3× faster in training than a Transformer while maintaining better perplexity — which directly determines the depth of the model's understanding of DNA.

---

4. A Three-Act Experiment: AI Brainstorms → Computer Filters → Reality Decides

Act I: AI Brainstorms (Generation)

Input: the 5′ end of ΦX174 as prompt (10 bp to 500 bp) Models: Evo 1 SFT (supervised fine-tuned) and Evo 2 SFT Operation: autoregressive generation, temperature sweep (0.7–1.2) Output: 1,000 sequences per configuration, thousands of candidate genomes in total

The key variable is generation temperature. Low temperature (0.7) is conservative, producing sequences closer to natural phages; high temperature (1.2) is bold and more divergent — but more likely to produce biologically infeasible sequences.

The team found medium temperature (~0.9) and medium prompt length (~100 bp) were the sweet spot: preserving genomic architectural plausibility (architecture similarity score > 0.38) while achieving high sequence novelty.

Act II: Computer Filtering (Computational Filtering)

From thousands of sequences, the team applied a six-layer filter framework:

1. Viral classification (geNomad): confirm the sequence is viral, not bacterial or noise 2. Novelty detection (nucleotide BLAST): low query coverage and similarity against training data (Microviridae) 3. Quality assessment (CheckV): >87% of sequences rated "High Quality" or "Complete" 4. Gene annotation (PHROGs database): predicted protein functions match known phage proteins 5. Structural plausibility (ESMFold): predicted pLDDT scores comparable to natural proteins 6. Architectural similarity (custom algorithm): compare genome architecture to ΦX174 to ensure sensible ORF layout

These six layers reduced thousands of sequences to 302 — each passing the computational test of "this looks like a plausible phage genome."

Act III: Reality Decides (Experimental Validation)

The 302 sequences were synthesized, cloned into plasmids, electroporated into E. coli C, and subjected to plaque assays.

Result: 16 formed plaques (5.3%).

But 5.3% was just the start. The 16 survivors faced stricter tests:

• Growth competition: several AI-designed phages showed higher fitness than natural template ΦX174
• Lysis kinetics: faster drop in bacterial OD600 after infection, indicating higher lysis efficiency
• Cryo-EM: the complete Evo-Φ36 virion was solved at 2.9 Å, proving AI-designed proteins assemble correctly
• Resistance overcoming: an AI-designed phage cocktail fully suppressed growth of ΦX174-resistant strains within 5 passages

---

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 biological complexity.

Bottleneck 1: A Genome Is a Coupled System, Not a Parts List

A phage genome is not a linear combination of 11 independent genes. Expression levels, translation efficiency, protein folding, and interactions are all 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 this means one gene's "error" can be offset by another gene's "adjustment," and such compensation cannot be fully predicted by computational screening.

Essence: this is an NP-hard combinatorial optimization problem. 302 candidates cover only a tiny corner of the design space.

Bottleneck 2: The Hard Ceiling of Experimental Throughput

302 designs already push high-throughput synthesis limits. Each design requires:

• Gene synthesis (~$100–500 per construct)
• Cloning, transformation, culture
• Plaque assays and confirmation

At a 5% success rate, finding one survivor requires testing ~20 designs on average — an expensive game of luck. That the team tested 302 to find 16 shows computational filtering works but still produces many false positives.

Future breakthroughs:

• Cell-free systems: rapid screening of genome activity in a test tube
• ML-guided experimental design: Bayesian optimization or active learning to pick the most informative candidates
• Higher-fidelity simulation: end-to-end from gene sequence to proteome to cellular dynamics

Bottleneck 3: Context-Dependent Structural Compatibility

Cryo-EM shows the N-terminal 14 amino acids of Evo-Φ36's J protein are unresolvable inside the capsid — a flexible/disordered segment, radically different from ΦX174's rigid binding mode.

This means Evo's designs are not "standard answers" but alternative solutions outside known constraints. That creativity is remarkable, but it implies:

• We cannot use traditional structural-biology intuition to predict which designs will succeed
• Each success is a small scientific discovery requiring experimental validation
• Confidence calibration of computational models remains an open problem

---

6. Biosafety vs. the Post-Antibiotic Era: Which Side Are You On?

6.1 Safety Guardrails

The Evo team was clear-eyed. Their safety design includes:

• Training data filtering: exclusion of 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: designs differ significantly from natural phages, lowering recombination risk with known pathogens
• Community norms: adherence to iGEM and DIY Bio safety guidelines

But these guardrails are imperfect:

• 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?

6.2 A Glimmer for the Post-Antibiotic Era

On the other hand, phage therapy is one of the most promising answers to antibiotic resistance:

• Phages are nature's bacterial predators, evolved over billions of years
• They are highly specific — infecting only particular bacterial strains, sparing the human microbiome
• But natural phage therapy suffers from narrow spectrum and resistance evolution

This work's breakthroughs:

• AI-designed phage cocktails overcome resistance: in experiments, bacteria failed to evolve resistance within 5 passages
• Programmable host range: host specificity can be tuned via spike protein sequence similarity
• Design space beyond natural diversity: AI can explore sequence space evolution never visited

6.3 Where Do I Stand?

In Feynman's spirit: 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 science — it's fear. The right questions are:

• How big is the risk?
• What experiments can quantify it?
• What are the benefits?

Antibiotic resistance kills 700,000 people globally per year. By 2050, that is projected to reach 10 million.

If an AI-designed phage can turn even one drug-resistant infection from "untreatable" into "treatable," the benefit deserves serious consideration.

But the precondition is: safety validation must come first. Not as an afterthought, not a one-liner in a paper — but like this work's cryo-EM: let the experimental data speak.

---

7. What Kind of Breakthrough Is This?

In the history of biology, this experiment sits roughly here:

• 1952: The Hershey–Chase experiment proves DNA is the genetic material
• 2003: Human Genome Project completed — we first *read* the book of life
• 2012: CRISPR-Cas9 — we learned to *edit* life
• 2026: Evo — we *wrote* an entirely new living system for the first time

Note the key words: entirely new and alive.

Previously we engineered viruses — inserting GFP into lentivirus, putting Cas9 into AAV. That was recombination, rearranging parts nature already had.

This was de novo design. Some genes in Evo's genomes have no significant homologs in natural databases. It is not 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 understanding structural constraints, but we have no mechanistic explanation — did it "know" the J protein needs to interact with the capsid somehow, or did it get lucky?

2. How high can the success rate go? 5.3% is stunning for basic research but far from enough for clinical use. A 100× to 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 131k context — but designing a functional 200 kb genome is a different challenge entirely.

4. What about eukaryotes? For safety, Evo's training excluded eukaryotic viruses. But phage-like therapy could theoretically work against eukaryotic pathogens (e.g., fungi). Should this path be pursued — and how, safely?

---

References

• 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