Omnii Phage

Antibiotic resistance is a growing global concern, contributing to nearly 5 million deaths per year¹. Phage therapy offers a promising alternative when antibiotics fail, but finding a phage that can kill a patient's specific bacterial infection remains a major challenge. We post-trained Omnii-1, our multimodal genome language model, to design phage variants to target specific bacterial strains. Here, we present Omnii-1's ability to generate viable phages, screen and prioritize variants, and predict phage-host interactions. In a world first, Omnii-1-designed phages were administered to a human patient, marking a major milestone toward personalized, AI-designed treatments for antibiotic-resistant infections.

Antibiotics have transformed medicine, making once-fatal bacterial infections routinely treatable. But bacteria evolve. Multidrug-resistant bacteria, often called superbugs, are making these drugs increasingly less effective. Through mutations and the spread of resistance genes, they develop resistance to antibiotic treatment. In 2021, bacterial antimicrobial resistance was associated with an estimated 4.7 million deaths worldwide.¹

One promising solution comes from bacteria's natural enemies: phages. These viruses are found throughout microbial ecosystems, from oceans and soil to the human gut.^2,3 They have evolved alongside bacteria for billions of years, developing ways to recognize and infect particular bacterial hosts. During infection, a phage attaches to a susceptible bacterium, delivers its genetic material, and replicates inside the cell. Eventually, the cell bursts, killing the bacterium and releasing new phages that can infect other bacteria.

The idea behind phage therapy is deceptively simple: find phages that can kill the bacteria causing a patient's infection and use them as a treatment. Phages do not infect human cells, and phage therapy is generally considered safe, though further clinical evidence and monitoring remain necessary.⁴ Unlike broad-spectrum antibiotics, phages can target particular bacteria with remarkable specificity, sparing beneficial bacteria such as those in the gut that support digestion. This specificity is both the promise and the challenge of phage therapy: treatments can be highly targeted, but suitable phages must be found or designed for each patient's infection.

Today, personalized phage therapy largely depends on empirical search, as the biological mechanisms governing phage-host specificity are not yet fully understood. Starting with a patient's bacterial isolate, researchers screen existing phage collections for effective candidates, turning to additional collections, environmental samples, or engineering when none are suitable. For patients with severe infections and few remaining options, these delays can be life-threatening. We need efficient ways to develop personalized phage therapies for patients who cannot afford to wait.

What if genome language models could design phages against the bacteria causing a patient's infection? We post-trained Omnii-1, our multimodal genome language model, to generate functional phage variants and predict phage-host interactions, evaluating these capabilities through wet-lab experiments and computational benchmarks.

This milestone marks an early step toward our broader goal of using AI to develop personalized treatments to save lives. Making phage therapy accessible to more patients will require robust workflows for designing effective phages against individual infections. Here, we describe our approach, the experimental results, and the challenges that remain.

Developing a personalized phage therapy begins with a bacterial isolate: a sample of the bacteria causing a patient's infection. The goal is to find or design phages that kill the target bacteria, with the broader potential to address emerging strains and restore susceptibility to antibiotics. Understanding what makes a phage effective against one bacterial strain but not another is central to this challenge.

Phage infection begins with molecular recognition. A phage must first attach to a compatible receptor on the bacterial surface, often with high specificity. These receptors can include proteins, lipopolysaccharides, capsules, and pili. Small differences in receptor structure can allow a phage to recognize one bacterial strain but not another.

Recognition, however, is only the first challenge. After attachment, a phage must deliver its genome, overcome bacterial defenses^‡‡ Restriction-modification systems cut phage DNA lacking protective methylation, while CRISPR-Cas systems use RNA guides to target matching phage sequences. Phages can counter these defenses through DNA modifications or proteins that inhibit restriction enzymes or CRISPR-Cas activity., and replicate inside the cell to produce new phages (Figure 2). Failure at any of these steps can prevent a phage from establishing a productive infection, even after successfully recognizing a bacterium.

These interactions make phage engineering a challenging biological design problem. Changes to a phage's genome can alter which bacteria it recognizes, but may also disrupt functions required for assembly or replication. Successful engineering requires improving activity against the target bacteria while preserving the machinery necessary for infection.

Researchers traditionally begin developing phage therapies by screening collections of naturally occurring phages against a patient's bacterial isolate. When suitable candidates cannot be found, engineering novel phage variants offers another possibility.

Designing a phage does not necessarily require generating an entirely new genome. In our initial experiments, we chose T7, a well-characterized phage that infects E. coli, and focused on its tail fiber, a protein involved in recognizing and attaching to bacterial surfaces (Figures 3–4). Changes to this protein can alter which bacterial strains a phage recognizes, making it a promising target for engineering host interactions. By modifying the DNA encoding this protein, we can create new variants of T7, each carrying a different sequence within the same broader phage genome. While several other phage components offer opportunities for engineering phage-host activity, this initial example focuses on modifying host recognition.

This approach allows us to redesign a biologically important component while preserving the broader machinery required for infection. Yet even a short DNA region has an enormous number of possible sequences, only a fraction of which may preserve phage function. Efficiently exploring this sequence space through random mutations and experimental screening is impractical.

Genome language models offer a way to navigate this space by learning patterns in biological sequences, generating promising variants for experimental testing, and guiding designs toward specific bacterial targets.

Recent advances in genomic machine learning have demonstrated the potential of AI-driven phage design. King et al. showed that Evo, a genomic language model, could generate viable ΦX174-like bacteriophages.⁵ Separately, Gaborieau et al. and subsequent studies demonstrated strain-level phage-host prediction from genomic information using traditional machine learning methods.^6,7,8,9 Together, these advances established two complementary capabilities: generating functional phage sequences and predicting which bacteria a phage can infect.

With Omnii-1, we aim to bring these capabilities together to tackle a more demanding challenge: designing phages with activity against a patient's specific infection. This first extends genomic design to more complex phage systems. We selected the E. coli-targeting T7 phage, whose 40-kilobase genome is more than seven times as large as ΦX174’s 5.4-kilobase genome (Figure 4).¹⁰ Beyond genome size, the central challenge is now engineering toward a desired therapeutic effect while preserving the machinery required for functionality.

Omnii-1 is pretrained on diverse genomic sequences in addition to multi-omics, learning patterns of DNA sequence and organization. We then further train it on phage genomes, adapting its knowledge to the biological systems we want to engineer.

Crucially, Omnii-1 is trained to generate phage DNA using context on both sides of a selected region. Unlike models that generate sequences only from left to right, Omnii-1 can design a region within an existing genome while accounting for the surrounding sequence. Much like editing a section of code, this allows us to modify the genetic instructions for a specific biological component while maintaining compatibility with the rest of the genome. This capability was especially important to our T7 experiments when we designed the DNA of the phage's tail fiber (Figures 3–4).

Generating functional DNA is only part of the challenge. A phage that can infect one bacterial host may still fail to do so in the target bacterium. Designing toward a particular host therefore requires understanding the interactions between phages and bacteria.

We train Omnii-1 on phage-host interaction data to learn representations of phages and bacteria that capture relationships associated with infection. These representations support predictions of which bacteria a phage may infect and provide a foundation for guiding variant design toward specific host interactions.

Our computational phage design workflow depends crucially on feedback from wet-lab experiments. We synthesize Omnii-1-designed DNA variants, incorporate them into phages, and screen their activity across panels of bacterial strains. These experiments reveal which designs are functional and how their activity varies across hosts, providing feedback for subsequent rounds of modeling and design.

Experimental throughput also informs our engineering strategy. Beyond enabling precise biological modifications, targeting selected genomic regions allows us to synthesize shorter DNA sequences, construct large variant libraries, and screen many candidates in parallel. This increases the number of designs we can evaluate while generating data to improve our models and reduces wet-lab latency.

Together, these capabilities bring Omnii-1 closer to making personalized phage design more efficient and scalable. We next highlight Omnii-1's ability to generate and prioritize viable phage variants, predict phage-host interactions, and contribute to a therapeutic phage cocktail that demonstrated in vitro activity against a patient's bacterial isolate before clinical administration.

A successful phage design campaign depends on two capabilities: generating viable variants across a broad mutational space, and screening those variants in silico to distinguish promising designs from less promising ones. The first lets us explore phages that differ from wild type and may exhibit new patterns of activity. The second helps reduce wet-lab turnaround time by prioritizing the strongest candidates for experimental characterization.

We first tested whether Omnii-1-selected designs could be recovered as phages after rebooting. A designed sequence must produce viable phages before its activity can be tested on a bacterial host. We measured recovery across increasing amino acid distance from wild type and compared selected designs with random amino acid changes.

Selected designs were recovered more often than random controls across the 1–30-change bins. At 11–15 amino acid changes, 69.9% of selected designs were recovered, compared with 4.8% of random controls (Figure 5A). These results show that Omnii-1 can identify changes compatible with phage viability well beyond small edits to wild type. No selected designs were recovered beyond 30 changes in this experiment.

We next measured activity on a bacterial host by sequencing phage populations before and after infection and comparing enrichment with wild type. In a broader design library, 2,161 of 6,694 variants with quantified enrichment exceeded wild type, including variants with as many as 28 amino acid changes. Enriched random controls had at most five changes (Figure 5B). The design library includes both predicted and previously measured designs and differs from the cohort used in the recovery experiment. Variants below measurement support are shown separately in the figure.

Together, these experiments show that selected phage designs can remain viable and retain activity across substantial sequence changes, expanding the range of variants available for further engineering.

A viable design pool provides a starting point; the next challenge is identifying variants with strong activity on a particular bacterial host. We evaluated whether an Omnii-1-based scorer could prioritize these variants using experimental measurements from two E. coli strains. For each strain, we ranked variants by predicted activity and compared the highest-ranked candidates with the full evaluated set. Predictions were made with mutation families held out from training.

For E. coli JW3601, the top 100 of 7,806 variants had a median measured activity of 13.4 sigma, compared with −0.3 sigma across the full set. Activity is expressed relative to the noise measured in synonymous wild-type controls. Of those 100 candidates, 97 exceeded the +2 sigma threshold, compared with 27.1% of the full set (Figure 7). For E. coli JW3594, 98 of the top 100 candidates exceeded the same threshold, compared with 28.0% of all 5,903 evaluated variants.

These results suggest that Omnii-1 can help researchers and clinicians select active variants from a large design library, focusing experimental testing on promising candidates and reducing wet-lab turnaround time.

An individual phage variant design campaign is only one part of the challenge. Making personalized phage therapy scalable requires a broader understanding of phage-host interactions. For a new patient's bacterial isolate, which phages should we start with? How can we guide designs toward that strain, understand the mechanisms governing infection, and select combinations of phages for effective therapeutic cocktails?

To address these questions, we post-trained Omnii-1 on experimentally measured phage-host interactions, learning representations of phage and bacterial genomes associated with infection. We next evaluate whether Omnii-1 can use these representations to predict interactions with previously unseen bacterial strains, providing a foundation for more host-informed phage therapy design.

Developing a personalized phage therapy begins with a patient's bacterial isolate, for which phage interactions may not yet have been measured. We evaluate whether Omnii-1 can use interaction data from other bacterial strains to predict which phages are most likely to infect a previously untested strain.

To test this, we use a dataset of experimentally measured phage-host interactions from Gaborieau et al.⁶ We hold out E. coli strains from interaction training while retaining the tested phages, mirroring the case of selecting promising candidates from an existing phage collection for a new patient isolate strain.

We compare Omnii-1 against two specialized machine learning methods for phage-host prediction: the method developed by Gaborieau et al.⁶ and GenoPHI.⁸ The method developed by Gaborieau et al. trains separate models for individual phages using features tailored to E. coli, while GenoPHI combines protein-family features with recursive feature elimination and ensemble learning. In contrast, we post-train Omnii-1 on phage-host interaction data using genomic contigs from phages and bacteria associated with infection, including receptor recognition and adsorption, without assuming the identity of the phage or bacterial species. We then predict infection using a linear layer on the resulting embeddings and test whether Omnii-1's genomic representations can capture patterns of phage-host compatibility competitive with methods developed specifically for this task.

On randomly held-out strains, Omnii-1 performs competitively with both specialized methods. Its top two recommendations include an experimentally active phage for 92.1% of eligible strains, compared with 84.2% for each baseline (Figure 8A). For a new bacterial isolate, this suggests that Omnii-1 could help identify promising phages with fewer experimental tests during therapeutic screening.

Randomly held-out strains, however, may have close relatives in the training data. A harder challenge arises when a patient's isolate belongs to a bacterial lineage poorly represented in existing interaction data. To test this, we hold out an entire E. coli clade comprising 37 strains. Omnii-1's ability to identify active phages among its top recommendations remains largely unchanged, while both specialized baselines decline more substantially (Figure 8B), demonstrating competitive performance even on more distantly related bacterial strains.

Together, these results suggest that Omnii-1's post-trained genomic representations can generalize phage-host predictions to unseen E. coli strains, including those from previously unrepresented lineages. We are excited to continue improving this ability to identify promising phages and better inform designs targeting unfamiliar bacterial hosts.

Beyond E. coli, our goal is to design phages against antibiotic-resistant infections caused by a broader range of bacterial species. One important target is Pseudomonas aeruginosa, which can cause persistent lung infections in people with cystic fibrosis and serious wound and burn infections. Antibiotic resistance makes these infections particularly difficult to treat.¹¹

Extending phage-host prediction to other bacterial species presents another challenge: experimental interaction data are often limited. Our P. aeruginosa dataset from Piya et al.¹² contains 7,308 phage-host interactions, compared with 32,681 E. coli interactions from Gaborieau et al.⁶ We asked whether Omnii-1 could improve predictions for P. aeruginosa by learning from both species. To test this, we compare models post-trained on P. aeruginosa phages and hosts alone with models post-trained on the combined E. coli and P. aeruginosa phage-host datasets, evaluating both on the same randomly held-out P. aeruginosa strains.

Trained on P. aeruginosa interactions alone, Omnii-1 performs competitively with GenoPHI⁸ on average precision (AP) and outperforms it on AUROC. Adding E. coli interactions, without additional P. aeruginosa measurements, further improves Omnii-1's performance, increasing its AP from 0.570 to 0.608 and surpassing GenoPHI (0.581) (Figure 9). In contrast, GenoPHI's performance declines with the additional cross-species data. These results suggest that general genomic foundation models like Omnii-1 may benefit from scaling interaction data, even across bacterial species, in ways that specialized machine learning approaches may not.

This is promising for pathogens with limited experimental data. As we expand our phage-host interaction datasets, we hope to build on this capability to improve predictions for less-studied pathogens, identify promising phages, and guide designs for patients whose infections fall outside well-characterized systems.

The capabilities explored above are steps toward making personalized phage therapy more efficient and accessible. Our efforts aim to translate advances in AI-guided phage design from computational modeling into clinical applications.

As noted earlier, Omnii-1-designed phages have now reached a human patient. Omnii-1-designed T7 phage variants were included in the first therapeutic cocktail containing AI-designed phages administered to a patient with an antibiotic-resistant E. coli infection. In laboratory testing, the cocktail eliminated the targeted E. coli isolate in vitro. Treatment proceeded under an FDA-authorized single-patient Investigational New Drug (IND) application through expanded access, also known as compassionate use.

Reaching this patient marks a major milestone, but making personalized phage therapy accessible to many more will require a reliable and scalable approach. Each new bacterial infection presents distinct challenges, demanding a deeper understanding of phage biology and more precise ways to engineer it.

We are working to bring these capabilities together into a systematic design process. Better models of phage-host interactions can help identify promising starting phages and guide designs toward specific bacterial strains. More efficient experimental workflows can validate designs, uncover new biology, and inform subsequent campaigns. Together with expanding clinical partners, we hope to make personalized phage therapies accessible to more patients.

Beyond therapeutic applications, we are excited by the opportunity to uncover fundamental principles of phage biology and translate that understanding into more precise genome design. This connection between biological discovery and therapeutic engineering is central to our work with Omnii-1.

Realizing this vision requires collaboration across AI, biology, and medicine. We welcome collaborations with experimental researchers, organizations with phage-host datasets, and clinical partners. We're also hiring researchers and engineers interested in biological modeling and therapeutic genome design.

Ultimately, we hope to turn advances in AI and biology into personalized therapies that save lives.