Genome-Scale Generative Design: What an AI Language Model Wrote, and Why It Worked 

Artificial intelligence has been used to help design biological parts for several years. Proteins, CRISPR-Cas complexes, transposable elements, promoters, and toxin-antitoxin pairs have all been generated by various AI models. However, designing a genome is a different problem, and AI-aided design at that scale has, up until now, been out of reach.  

Recently, researchers at Stanford University and the Arc Institute report crossing that line. Using the genome language models Evo 1 and Evo 2, they generated complete bacteriophage genomes, chemically synthesized nearly 300 of them, and recovered 16 viable phages.  

A cocktail of those designs successfully inhibited bacterial strains that had evolved to resist the template phage. A comparable cocktail of natural phages could not achieve this. 

Less than 1% of genomic research has used AI 

To place these results within a wider context, we analyzed an 11,259-document corpus retrieved from the CAS Content Collection™ covering AI/ML topics applied to genomes, spanning 1989 to 2026.  The corpus is dominated by predictive and clinical genomics. Work that actually uses AI to design genomic sequences is rare, accounting for only 88 documents or 0.78% of this corpus. 

The 88 break down as follows: 

• CRISPR and editing design (guide RNA selection, off-target and editing-outcome prediction) is the most prevalent topic with 42 documents; it’s also the most established, having run continuously from 2018. 

• Regulatory element design (promoters, enhancers, UTRs) accounts for 18 documents. 

• Generative sequence design,(models that emit new nucleotide sequence rather than score existing sequence) accounts for 10 documents and is the newest, with a median publication year of 2025. 

• Whole-genome and chromosome-scale design account for eight documents. 

These are small numbers, and the newest categories are the smallest, so they indicate where activity is starting rather than a growth rate. Of the 88 design documents, 74 are journal articles, and only eight are patents: generative genome design is still academic literature. 

How many “genomic language models” are there?

Genome language models, however, are a crowded field. A 2026 systematic review screened 469 records and identified 58 model-development studies. Most are encoders – put a sequence in, and a score or an embedding comes out. These dominate the corpus. DNABERT and DNABERT-2 appear in 34 documents, Enformer in 13, Caduceus in five, HyenaDNA and the Nucleotide Transformer in four each. Yet, only three of those documents fall in our verified genome-design set. AlphaGenome, the strongest recent model of this type, reads up to a million bases and predicts regulatory consequences, but does not write sequences. 

Besides published models, there are a few under development that generate sequences available only as preprints with code and weights posted publicly (see Table 1). 

Table 1. List of genome LLMs capable of designing genome 

Commercial activity currently sits around the edges of this topic rather than at its center.  

• Profluent, which reports designing a CRISPR system entirely with AI, has announced a partnership with Eli Lilly on AI-designed recombinases for genetic medicine.  

• Dyno Therapeutics designs AAV capsids for gene delivery and has licensed them to Astellas and Roche. 

• Ginkgo Bioworks and Basecamp Research supply training data rather than models.  

• NVIDIA co-developed and distributes Evo 2 with the Arc Institute. 

• Tatta Bio, a nonprofit, has released the mixed-modality model gLM2.  

Programs aimed at designing genomes themselves remain scarce, and the genome-scale work so far has come largely from nonprofit and academic groups with compute partners (consistent with the patent picture in our corpus). 

Within that group, Evo has a longer track record and larger context. Evo 1 already produced functional CRISPR-Cas and transposon systems and megabase-scale sequences of plausible genomic architecture, and Evo 2 reads and writes at single-nucleotide resolution across a 1-million-token window. That context is the decisive property: a 5.4-kilobase phage genome fits inside it whole, so the model holds gene order, regulatory placement, and packaging signals in view at once rather than assembling a genome from independently scored fragments. The phage-heavy pretraining matters too, more than 2 million bacteriophage genomes, placing the target inside the distribution the models learned. 

What were the models asked to do? 

The template was ΦX174: a small lytic member of the Microviridae, whose 5.4-kilobase genome carries 11 genes and at least seven regulatory elements. Compactness, decades of characterization, and a harmless laboratory host made it an unusually safe proving ground. Both models were fine-tuned on approximately 15,000 Microviridae sequences before any candidate went to synthesis. 

AI as one step in a six-stage pipeline  

Raw generation did not produce viable genomes. The difference between this attempt and a failed one lies largely in what surrounded the model. The authors describe a six-stage pipeline:  

1. Pretraining

2. Supervised fine-tuning on Microviridae

3. Prompt engineering with ΦX174 sequence

4. Inference-time steering by separate models that scored genomic architecture and predicted host range

5. Computational filtering

6. Experimental validation

Prompting alone was not enough: at no prompt length did the base models recover ΦX174, whereas the fine-tuned versions did so. 

Filtering was where most candidates died in three tiers: basic sequence quality, specificity for the intended host, and distance from natural sequence. Quality control enforced lengths of four to six kilobases, GC content between 30 and 65%, and no homopolymer longer than 10 bases. The harder checks were biological, and this is where structure and function prediction entered the design loop rather than the write-up. Generated proteins were folded with ESMFold, and their mean pLDDT scores matched natural proteins while exceeding scrambled controls, which was evidence that the sequences encoded foldable proteins, not plausible-looking nucleotide strings. Genomes were scored by the viral classifier geNomad, and their predicted proteins aligned against the OpenGenome and PHROGs databases, returning a realistic distribution of functional annotations despite low sequence identity to anything known.  

None of six widely used gene annotation tools could annotate all 11 genes of wild-type ΦX174, because the genome's overlapping reading frames defeat standard open-reading-frame prediction, so the team built a bespoke coding-sequence predictor. Without it, there was no way to confirm that a generated genome carried a complete gene set - a generate-and-filter strategy with no filter. 

AlphaFold 3 appears as well, but downstream of design: predicting the orientation of the unusual J protein in the Evo-Φ36 capsid, and placing resistance-associated mutations on the virion exterior. Structure prediction played two roles: a filter during design and an explanatory tool afterward. 

Of 302 curated designs, 285 were successfully synthesized and assembled. The remainder failed on synthesis complexity. Sixteen inhibited the growth of Escherichia coli C - 12 of 227 sequences from Evo 1 (5.3%) and 4 of 58 from Evo 2 (6.9%). Viability tracked sequence similarity to natural genomes, reaching 46.2% among the closest designs: the further the model ventures from what evolution has sampled, the more often it fails. 

Three reasons this is a genuine advance 

1. Scale. Every prior generative success in this lineage was a part. This is a complete, self-replicating genome whose components must function together in a living host, delivered through an end-to-end framework rather than a single lucky construct.

2. The designs are not recombinations of the template. Cryo-electron microscopy of Evo-Φ36 (one of the 285 synthesized AI-designed phage) at 2.9 Å resolution showed that the designed phage carries a truncated DNA packaging protein J from Escherichia phage G4, a distant relative (25 residues in place of 38). That swap had been reported as nonviable in wild-type ΦX174. The model assembled a context in which an otherwise incompatible part functions with this being the clearest evidence that it internalized epistatic constraints rather than memorized sequence.

3. The designs carry real novelty. Mutations with no natural counterpart, divergent genes, altered regulatory sequence, and variable genome lengths. Editing clustered in promoter A, gene J, and terminator J: all regulatory and structural control points. 

Outpacing bacterial resistance

Phage therapy has drawn sustained interest as antimicrobial resistance has deepenedThe World Health Organization ranks antimicrobial resistance (AMR) among the top global health threats but remains limited by poor in vivo efficacy and by bacteria evolving resistance to the phages. 

Remarkably, a mixture of the AI-designed phages reported by King et al. were effective against ΦX174-resistant E. coli strains while a comparable mixture of natural ΦX174-like phages was not. 

Distance to the clinic 

Clinical translation remains early with ~96 interventional studies involving phage-based agents being listed on ClinicalTrials.gov (as of 12 August 2026) with nearly half of them recruiting, active, or not yet started. Among trials reporting a phase, Phase 1 and Phase 2 dominate (62 of 71). Targets include Pseudomonas aeruginosa in cystic fibrosis, urinary tract infections, diabetic foot ulcers, and prosthetic joint infections. Engineered phages have already reached patients: SNIPR001 is in a Phase 1b/2a study (NCT06938867) in stem-cell transplant patients carrying fluoroquinolone-resistant E. coli. 

The challenges of designing at this scale 

The designed phages were tested in E. coli C, a harmless laboratory strain, not a clinical pathogen. Phages aimed at Klebsiella or Pseudomonas typically run 40 to 100 kilobases, an order of magnitude larger than the 5.4-kilobase template used here, and the authors identify the cost of synthesizing and assembling that much DNA as the principal obstacle to designing them. The familiar hurdles for any biological medicine (manufacturing consistency, immunogenicity, endotoxin control) are unchanged by an AI-designed genome. 

Computing power is the obvious place to expect a wall, but that’s not the case, here. Evo does not use the standard transformer architecture, whose cost grows with the square of sequence length. Rather, its cost rises closer to linearly, so a 100-kilobase phage, or even a 4.6-megabase bacterial genome, is a manageable step up. Nor was context length stretched here: the models read eight kilobases at a time, comfortably more than the genome they were writing. A 3.1-gigabase human genome is a different matter and would require architectural work. What scales badly is viability. Of 285 genomes successfully synthesized, 16 produced working phages (5.6%), rising to 46.2% only among the designs closest to natural sequence. A single mutation anywhere can kill a genome, and every additional kilobase adds genes, regulatory elements, and packaging signals that must each be correct and mutually compatible. The inference is that the burden growing steeply with genome length is experimental, not computational. The number of designs that must be built and screened to recover a working one rises far faster than the compute needed to write them. 

Training data availability sets a second limit. The result rests on more than 2 million bacteriophage genomes in pretraining and roughly 15,000 close relatives of the template in fine-tuning, so the target sat well inside the distribution the models had learned. For a bespoke 100-kilobase phage against a specific clinical isolate, that density of near neighbors is not available, and for eukaryotic genomes the problem changes in kind. Function there depends on chromatin state, splicing, and long-range regulation that primary sequence underdetermines. The wet-lab precedent is sobering. A 1.08-megabase bacterial genome was chemically synthesized and transplanted in 2010 and a minimized successor followed in 2016, yet a fully consolidated synthetic yeast strain remains unfinished after roughly two decades. Enabling methods are advancing, but none yet offers the throughput that made a 285-design screen possible at phage scale.  

The authors also recommend involving safety and security specialists at every stage, advice that carries more weight as targets move from a harmless laboratory phage toward clinical pathogens. 

The result is narrow and solid: given a well-characterized template, a language model wrote genomes that became working viruses and beat natural ones at breaking resistance. Read against the wider literature, it's a major marker in a field that spent years perfecting guide RNA selection and has only recently begun writing sequence outright. Scaling to a therapeutic phage is now as much a synthesis and safety problem as a modeling one.