Radical Numerics creates first AI-generated phages to reach human patient with Omnii-1

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. Understanding the genetic features that shape a phage's specificity for particular bacterial strains 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.

However, recognition is only the first challenge. After attachment, a phage must deliver its genome, overcome bacterial defenses^bb For example, restriction-modification systems cut phage DNA lacking protective methylation, while CRISPR-Cas systems use RNA guides to target matching sequences. Phages can counter these defenses through DNA modifications or proteins that inhibit these systems., 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 are unavailable, engineering phage variants offers another possibility. Previous clinical applications have included engineered phages modified to promote lytic infection or equipped with CRISPR systems to enhance antibacterial activity^5,6.

Similarly, using AI to generate new phage variants does not require designing an entirely new genome. Instead, we can modify selected regions involved in particular biological functions. For example, receptor-binding proteins help determine which bacterial strains a phage can recognize by binding to specific receptors on the bacterial surface, making them promising targets for engineering host specificity. By changing the DNA encoding these proteins, we can explore new patterns of host recognition while preserving the broader phage genome.

Still, even a short DNA region has an enormous number of possible sequences^cc A region of n nucleotides has 4ⁿ possible DNA sequences., only a fraction of which may preserve phage function. Exploring this sequence space through random mutations and experimental screening alone is resource intensive at scale. Genome language models offer a way to generate promising variants, guide designs toward specific bacterial targets, and prioritize them for experimental testing.

Recent advances in machine learning have demonstrated complementary capabilities for phage genome design and phage-host interaction prediction. King et al. showed that Evo, a genome language model, could generate viable ΦX174-like bacteriophages (hereafter “EvoPhage”)⁷. Separately, Gaborieau et al. and subsequent studies demonstrated strain-level phage-host prediction from genomic information using machine learning methods^8,9,10,11.

Our goal is to build on these capabilities to design phages with therapeutic activity against specific bacterial strains.

Omnii-1 is pretrained on diverse genomic sequences and multiomics data, then further trained on phage genomes to adapt its knowledge to these biological systems.

Among the phages we have engineered with Omnii-1 is T7, a well-characterized E. coli phage¹² that we use here to illustrate our approach to genome design. T7 has an approximately 40-kilobase genome, more than seven times the size of the 5.4-kilobase ΦX174 genome used in EvoPhage (Figure 4)¹³.

Importantly, to engineer specific biological functions while preserving phage viability, Omnii-1 can generate DNA using context on both sides of a selected design region, accounting for its surrounding genomic sequence. In our T7 experiments, we applied this capability to gp17, T7’s receptor-binding tail fiber protein (Figure 3).

Generating functional phages is only part of the challenge. A phage that infects a laboratory strain may not infect a patient's bacterial isolate, making it essential to understand phage-host compatibility.

We trained Omnii-1 on publicly available datasets of experimentally measured phage-host interactions to model compatibility between phages and bacterial strains. These predictions can help identify candidate phages for engineering or therapeutic cocktails, while the learned representations provide a foundation for guiding designs toward specific bacterial targets.

We also use activity measurements from our design campaigns to predict the fitness of engineered variants against particular bacterial strains with Omnii-1. These finer-grained predictions draw on Omnii-1's understanding of phage biology and host interactions to prioritize designs for experimental testing.

Experimental feedback is essential to refining these capabilities. We synthesize Omnii-1-designed phage variants and measure their recovery and activity across bacterial strains. These experiments help characterize the activity of our variants against target hosts and inform subsequent rounds of modeling and design.

We demonstrate two capabilities important to our phage design campaigns: generating viable designs across mutational space and prioritizing variants with stronger activity against specific bacterial strains. We illustrate both using an example T7 tail fiber design round.

We tested whether Omnii-1-designed variants could produce viable phages after rebooting^dd “Rebooting” introduces an assembled phage genome into a permissive bacterial host to produce phages. Recovery, the fraction of attempted designs detected with sufficient sequencing support, is a proxy for viability. Undetected designs are not necessarily nonviable. by measuring recovery across increasing numbers of amino acid changes from wild type. At 11–15 amino acid changes, 69.9% of designed variants were recovered, compared with 4.8% of random controls (Figure 5A). This advantage held across groups spanning 1–30 amino acid changes from wild type, showing that Omnii-1 can identify sequence changes compatible with phage viability even at substantial distances from wild type.

Recovery alone does not establish whether phage variants retain efficacy across mutational space. We therefore assessed activity following infection of a bacterial host, using enrichment relative to wild type as our measure. Of 6,694 variants with measurable activity, 2,161 exceeded wild type, including variants with up to 28 amino acid changes. Random controls that exceeded wild type had at most five amino acid changes (Figure 5B). These results suggest that Omnii-1 can identify phage variants that remain active despite substantial sequence changes.

We next evaluated whether Omnii-1 could prioritize variants with higher activity against specific bacterial strains. Using experimental measurements from our design campaigns, we ranked variants held out from training by predicted fitness and compared their measured activity with the full evaluated library across three bacterial strains (Figure 6).

For two strains, the library exhibited a broad range of activity, from low to high. Omnii-1 prioritized higher-activity designs for both, with 97 and 96 of its top 100 candidates, respectively, exceeding the activity threshold. The rankings shared 57 of their top 100 variants while also identifying distinct candidates for each strain.

The third strain presented a more challenging activity landscape, with only 4.5% of evaluated variants exceeding the threshold. Among Omnii-1's top 100 candidates, 24% exceeded that threshold, more than five times the proportion in the full library. Its highest-ranked candidates did not overlap with those selected for the other two strains (Figure 6B). These results highlight the value of strain-specific prioritization when selecting promising designs for experimental testing, particularly when higher-activity variants are scarce in the initial library.

Beyond individual phage design campaigns, developing personalized therapies at scale requires understanding phage-host interactions across diverse bacterial strains. We examine how well Omnii-1's predictions generalize to bacterial hosts not represented in its training data.

We trained and evaluated Omnii-1 using publicly available phage-host interaction data from Gaborieau et al.⁸ (hereafter “Gaborieau”). We held out E. coli strains from interaction training while retaining the tested phages, assessing whether Omnii-1 could predict interactions with unseen hosts using measurements from other strains. This mirrors the challenge of identifying promising phages from an existing collection for a new patient's bacterial isolate.

We compared Omnii-1 against two specialized machine learning workflows for phage-host prediction: Gaborieau and GenoPHI¹⁰. Gaborieau 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. Omnii-1 uses genomic contigs from phages and bacteria without requiring species-specific engineered features. We tested whether its genomic representations could support phage-host predictions competitive with methods developed specifically for this task.

On randomly held-out strains, Omnii-1 performed competitively with both specialized workflows (Figure 7A). However, these strains may have close relatives in the training data. We therefore tested generalization to more distantly related hosts by holding out an entire E. coli clade comprising 37 strains. Omnii-1 maintained higher rates of identifying active phages than both specialized workflows across the top five recommendations, reaching 91.4% of eligible strains with just two phages (Figure 7B). These results suggest that its predictions remain useful even for bacterial lineages absent from interaction training.

Beyond E. coli, designing phages against antibiotic-resistant infections caused by a broader range of bacterial species is essential. One important target is Pseudomonas aeruginosa, which can cause serious infections in patients with wounds or burns, as well as persistent lung infections in people with cystic fibrosis. Antibiotic resistance can make these infections particularly difficult to treat¹⁴.

Extending phage-host prediction to other bacterial species presents a challenge: experimental interaction data are often limited. The P. aeruginosa dataset from Piya et al.¹⁵ contains 7,308 phage-host interactions, compared with 32,681 E. coli interactions from Gaborieau⁸. We asked whether Omnii-1 could improve predictions for P. aeruginosa by learning across species. We compared models trained on P. aeruginosa interactions alone versus combined E. coli and P. aeruginosa interaction data, evaluating both on the same randomly held-out P. aeruginosa strains.

With P. aeruginosa interaction data alone, Omnii-1 performed competitively with GenoPHI¹⁰ on average precision (AP) and outperformed it on AUROC. Adding E. coli interactions, without additional P. aeruginosa measurements, further improved Omnii-1's performance, increasing its AP from 0.570 to 0.608 and surpassing GenoPHI (0.581). GenoPHI's performance, by comparison, declined with the additional cross-species data (Figure 8). These results suggest that Omnii-1 can benefit from scaling interaction data across species, even without additional measurements from the target host.

This ability to learn across species is particularly promising for pathogens with limited experimental data. As we expand our phage-host interaction datasets, we hope to improve predictions for less-studied pathogens and guide designs for infections outside well-characterized systems.

Bridging current gaps in phage therapy design with Omnii-1 is ultimately aimed at enabling reliable, effective phage therapies for patients in need. In the world's first clinical use of AI-designed phages, designs generated by Omnii-1 were administered to a patient with an antibiotic-resistant bacterial infection under an authorized, physician-led compassionate-use pathway.

In laboratory testing, the cocktail containing Omnii-1-designed phage variants eliminated the patient's antibiotic-resistant isolate in vitro. The investigational cocktail was subsequently administered to the patient under compassionate use.

Compassionate use allows patients with serious conditions to receive investigational therapies outside clinical trials when satisfactory alternatives are unavailable, subject to oversight and safety requirements established by the relevant government regulatory authorities. To protect the patient's privacy, we are not disclosing details of their clinical course or outcomes, as even limited clinical information could make the patient identifiable. These investigational therapies have not been approved by regulatory bodies for broader clinical use, and observations from a single patient cannot establish general safety or efficacy.

This marks an important milestone in translating AI-designed biology from computational predictions and laboratory validation to clinical administration. Making personalized phage therapy accessible to more patients will require a reliable and scalable approach. Each new bacterial infection presents distinct challenges, demanding a deeper understanding of phage biology and more precise approaches to phage engineering.

Our goal is to develop a systematic process for designing phages against individual bacterial infections. Better models of phage-host interactions can help identify promising starting phages and guide designs toward specific bacterial strains, while more efficient experimental workflows can validate designs, uncover new biology, and inform subsequent campaigns.

Alongside this work, we are committed to rigorous precautions and procedures for safe wet-lab research. Working with others in this field, we will continue to help shape protocols and norms of responsible governance, while adapting our biosecurity, biosafety, containment, disposal, and data-sharing measures commensurate with advances in Omnii and other AI models.

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 and computational biologists, organizations with phage-host datasets, and clinical partners. Clinical partners can provide bacterial isolates for screening against Omnii-1-designed phages. When promising candidates are identified and validated, we aim to work with treatment teams to manufacture patient-specific phage doses for potential use under appropriate regulatory pathways, including compassionate use.

We are also hiring AI researchers, engineers, and biologists interested in advancing biological modeling and therapeutic genome design.

We hope to turn advances in AI and biology into personalized therapies that save lives.

We're hiring → join us