AI-Designed Phages Reach a Human Patient: Radical Numerics Unveils Omnii-1
On October 20, 2026, American biotechnology company Radical Numerics announced that bacteriophages designed using its Omnii-1 artificial intelligence model had been administered to a human patient with an antibiotic-resistant Escherichia coli infection. According to the company, the investigational phage cocktail included AI-designed variants of bacteriophage T7 and was administered under a single-patient expanded-access authorization from the U.S. Food and Drug Administration (FDA).
The announcement represents a potentially important development at the intersection of generative artificial intelligence, synthetic biology and personalized phage therapy. While genome language models have already demonstrated their ability to design functional bacteriophages in laboratory experiments, Radical Numerics reports taking a further step by incorporating AI-engineered viral variants into a preparation administered to a patient.
However, the distinction between technological achievement and therapeutic efficacy is essential. Radical Numerics has not disclosed the patient's clinical outcome, citing privacy concerns. The company reports that the investigational cocktail eliminated the patient's bacterial isolate under laboratory conditions, but has not released sufficient clinical information to establish whether the treatment resulted in bacterial eradication, clinical improvement or recovery.
The significance of Omnii-1 therefore lies in the reported transition from computational genome design to experimental validation and authorized human administration, rather than in any demonstrated clinical success.
From Natural Phage Discovery to AI-Guided Genome Engineering
Personalized phage therapy traditionally begins with the isolation of the bacterium responsible for a patient's infection. Researchers then screen collections of naturally occurring bacteriophages to identify viruses capable of infecting the bacterial strain. This process can become particularly challenging when the pathogen exhibits substantial genetic diversity or carries resistance mechanisms preventing infection by available phages. Even closely related E. coli strains may display markedly different susceptibility profiles because of variations in surface receptors, lipopolysaccharides and intracellular antiviral defenses.
Radical Numerics is developing Omnii-1 to address this problem through both phage–host interaction prediction and genomic engineering. The model is trained on genomic and multiomics information, then further adapted to bacteriophage sequences and experimentally measured interactions. Unlike tools that only rank existing phages, Omnii-1 can generate modifications within selected regions of a viral genome while accounting for surrounding genetic sequences. The aim is to identify mutations that alter useful biological properties without compromising the phage's ability to reproduce.
The company illustrates this approach using bacteriophage T7, a well-characterized virus infecting E. coli with a double-stranded DNA genome of approximately 40 kilobases. Its experiments focus particularly on gp17, the tail fiber protein involved in bacterial receptor recognition. By modifying the DNA encoding this protein, researchers can explore different patterns of bacterial recognition while preserving essential components of the viral replication machinery.
This distinction is important because Omnii-1 does not necessarily design an entirely new bacteriophage genome from scratch. Its demonstrated T7 engineering strategy modifies selected genomic regions within an existing viral framework. Such targeted engineering is particularly attractive for therapeutic development because an otherwise functional bacteriophage may acquire useful changes in host recognition without requiring the redesign of every biological process involved in infection.
Experimental Results: AI-Designed Phages Outperform Random Mutations
Radical Numerics has released experimental data supporting Omnii-1's ability to generate viable phage variants across substantial genetic distances from their parental sequence. The researchers compared computationally selected designs with randomly generated mutations and evaluated whether the resulting genomes could produce recoverable bacteriophages after reconstruction in bacterial cells.
Among variants containing 11 to 15 amino acid changes, 69.9% of Omnii-1-designed sequences were recovered, compared with only 4.8% of random controls. This approximately 14.6-fold difference suggests that the model can distinguish sequence combinations compatible with phage viability more effectively than random mutagenesis in the tested experimental system. However, recovery was assessed using sequencing-based evidence and should not be interpreted as a direct measurement of therapeutic effectiveness.
The researchers also evaluated the biological activity of 6,694 variants for which enrichment measurements were available. Of these, 2,161 demonstrated greater enrichment than the wild-type phage, including some containing as many as 28 amino acid substitutions. Enrichment reflects relative performance under the experimental selection conditions and is useful for identifying promising variants, but it does not automatically indicate equivalent improvements in bacterial killing or clinical efficacy.
A further experiment investigated whether Omnii-1 could prioritize high-activity candidates against particular bacterial strains. For two tested strains, 97 and 96 of the model's top 100 recommendations, respectively, exceeded the predefined activity threshold. In a more challenging third strain, where only 4.5% of the evaluated variants exceeded that threshold, 24% of the top 100 AI-selected variants were active according to the same criterion.
These results indicate that the model can enrich a candidate selection for variants exhibiting useful experimental activity, even when active sequences are relatively uncommon. This could reduce the burden of screening large engineered phage libraries, although the degree of time saving, reproducibility between laboratories and performance against clinically diverse bacterial isolates require further assessment.
Predicting Which Phages Can Infect Previously Unseen Bacterial Strains
Beyond generating new variants, Omnii-1 incorporates a phage–host interaction prediction capability intended to identify which bacteriophages are most likely to infect a bacterial strain. This function is particularly important for personalized medicine, where a patient's isolate may differ genetically from the bacterial strains represented in existing phage collections.
Radical Numerics evaluated the model using published E. coli phage interaction datasets and compared it with specialized prediction methods, including GenoPHI and the approach developed by Gaborieau and colleagues. To assess generalization, the company excluded an entire evolutionary group containing 37 E. coli strains from interaction training, creating a more demanding evaluation than randomly withholding closely related isolates.
Under these conditions, Omnii-1 identified at least one active phage among its first two recommendations for 91.4% of eligible bacterial strains. This is a measure of successful candidate identification within the evaluation dataset, not a 91.4% probability that any newly designed phage will infect an arbitrary bacterial strain.
The company also investigated whether genomic information from one bacterial species could improve predictions for another. Incorporating E. coli interaction data into training for Pseudomonas aeruginosa increased Omnii-1's average precision from 0.570 to 0.608, exceeding the reported GenoPHI benchmark of 0.581 under the corresponding evaluation conditions.
These findings suggest that genomic language models may learn certain transferable features of phage–bacteria interactions. Such transfer could be particularly useful for bacterial pathogens with limited experimentally characterized phage collections. Nevertheless, the reported benchmarks were produced by the developer, and broader independent evaluations will be necessary to establish how reliably these capabilities generalize across diverse bacterial species and phage families.
An FDA-Authorized Compassionate-Use Treatment for One Patient
The most consequential element of the October announcement is the reported administration of an investigational phage cocktail containing Omnii-1-designed T7 variants to a patient infected with antibiotic-resistant E. coli.
According to Radical Numerics, laboratory testing showed that the cocktail eliminated the patient's bacterial isolate in vitro before the preparation was administered. The treatment proceeded through an FDA-authorized single-patient Investigational New Drug application, under the expanded-access pathway and physician supervision.
The FDA's expanded-access framework allows investigational treatments to be provided to individual patients with serious or life-threatening conditions when satisfactory alternatives are unavailable and applicable requirements are met. Such authorization permits a specific investigational use; it does not establish that the treatment has been approved for routine clinical practice.
The company has not publicly disclosed the infection's anatomical location, the complete phage cocktail composition, the administered dose, treatment duration, route of administration, microbiological follow-up or clinical outcome. Radical Numerics attributes this decision to patient confidentiality, noting that even limited clinical details could potentially permit identification.
Respecting that confidentiality is appropriate, but the absence of outcome data substantially limits clinical interpretation. Laboratory elimination of an isolate does not demonstrate that the same phages achieved bacterial eradication inside the patient, where immune responses, tissue distribution, bacterial physiology and the emergence of resistance can influence treatment activity.
The announcement therefore supports the reported feasibility of incorporating AI-designed phage variants into a medically supervised investigational treatment. It does not yet establish clinical effectiveness, general safety or superiority over conventional phage selection and engineering approaches.
How Does Omnii-1 Compare With Previous Engineered Phage Therapies?
Although Radical Numerics describes the administration as a world first involving AI-generated bacteriophages, engineered phages have previously been administered to human patients.
In 2019, researchers led by Graham Hatfull and collaborators at the University of Pittsburgh, working with clinicians at Great Ormond Street Hospital, reported the treatment of a teenager with disseminated drug-resistant Mycobacterium abscessus infection using a three-phage cocktail that included genetically engineered variants. The treatment was associated with substantial clinical improvement, establishing an important precedent for therapeutic phage engineering.
More recently, Samuel King, Brian Hie and colleagues from Stanford University and the Arc Institute demonstrated the generation of complete functional bacteriophage genomes using the Evo family of genomic language models. Their research, published in Science in August 2026, yielded 16 experimentally viable phages related to ΦX174 and showed that cocktails of generated variants could overcome resistance in laboratory E. coli strains.
The distinction in the Radical Numerics announcement is therefore not the first use of engineered phages, nor the first experimental demonstration of AI-generated bacteriophages. It is the company's reported clinical administration of bacteriophage variants designed using a generative AI model. That narrower milestone is significant, although its claimed worldwide priority has not been independently established through publicly available clinical documentation.
The announcement also comes amid growing interest in AI-driven phage research. On the same day, October 20, French biotechnology company Phagos released CAPHARD, a 900-million-parameter model designed to predict phage–bacteria interactions from genomic sequences. CAPHARD primarily helps researchers rank existing candidates, whereas Omnii-1 additionally demonstrates the generation and experimental prioritization of engineered viral variants. The two systems therefore address related but distinct stages of computational phage discovery and design.
Radical Numerics and the Growing Industry of Generative Biology
Radical Numerics was established by researchers with experience developing large genomic foundation models, including Evo and Evo 2. In June 2026, the company announced $50 million in seed financing to advance its work in biological artificial intelligence, therapeutic engineering and biosecurity.
Omnii-1 forms part of a broader research platform designed to integrate DNA sequences with experimental biological measurements. According to the company, the model was pretrained using 28 biological modalities, including information about DNA, RNA, proteins, gene expression, chromatin accessibility and epigenetic regulation. Its intended applications extend beyond phage engineering to genomic interpretation, molecular design and biological defense.
The phage programme represents an early attempt to translate these capabilities into therapeutic development. Radical Numerics is seeking collaborations with microbiology laboratories, organizations maintaining phage–host interaction datasets, hospitals and clinicians treating antibiotic-resistant infections. The company intends to combine computational design with experimental validation and, where appropriate, patient-specific manufacturing under applicable regulatory pathways.
As with other emerging AI-driven biotechnology platforms, the key question will be whether experimental successes can be reproduced systematically across different bacterial pathogens and translated into treatments meeting pharmaceutical quality, safety and efficacy requirements. The company's publicly available research preview provides substantial technical information, but it is not equivalent to an independently reviewed clinical study.
Toward Personalized Phage Therapy Designed With Artificial Intelligence
The Omnii-1 announcement illustrates an emerging development in bacteriophage research: moving beyond computational identification of naturally occurring viruses toward the generation of engineered variants tailored to particular bacterial targets.
The experimental results suggest that genomic language models can help identify viable mutations and prioritize promising phages more efficiently than random sequence selection in the evaluated systems. The reported clinical administration adds an important translational dimension, showing how these approaches may begin to enter medically supervised treatment pathways.
However, the available evidence supports a technological milestone rather than a therapeutic breakthrough. The patient's clinical outcome remains undisclosed, the FDA authorization concerns one investigational use, and the model's reported performance has yet to be established through comprehensive independent evaluation.
For personalized phage therapy, the longer-term opportunity is substantial: computational models could eventually connect bacterial genome sequencing, phage engineering, experimental susceptibility testing and treatment preparation within a more systematic development process.
Radical Numerics has reported an important step toward that objective. Whether AI-designed phages can consistently improve patient outcomes remains the question that future clinical research must answer.
Sources :
Radical Numerics — Iyengar A., Siranosian B., Sarkar A., Noonan A., Poli M., Nguyen E., et al. (October 20, 2026). Radical Numerics Creates First AI-Generated Phages to Reach Human Patient With Omnii-1. Primary company research announcement detailing genomic engineering experiments, prediction benchmarks and the reported compassionate-use administration. Original announcement.
Radical Numerics (October 20, 2026). Omnii Phage: Designing Targeted Phage Therapies. Supplementary company presentation identifying the antibiotic-resistant E. coli infection, T7 variants and single-patient FDA expanded-access authorization. Omnii Phage.
Radical Numerics (June 15, 2026). Radical Numerics Raises $50M Seed to Build General Biological Intelligence. Corporate background and financing announcement. Company announcement.
King S. H., Driscoll C. L., Li D. B., et al. (2026). Generative Design of Bacteriophages With Genome Language Models. Science, 393, eaec2657. Experimental generation of functional ΦX174-like bacteriophages using genomic language models. https://doi.org/10.1126/science.aec2657.
Dedrick R. M., Guerrero-Bustamante C. A., Garlena R. A., et al. (2019). Engineered Bacteriophages for Treatment of a Patient With a Disseminated Drug-Resistant Mycobacterium abscessus. Nature Medicine, 25, 730–733. An earlier clinical precedent for genetically engineered bacteriophages. https://doi.org/10.1038/s41591-019-0437-z.
U.S. Food and Drug Administration. For Physicians: How to Request Single Patient Expanded Access (Compassionate Use). Official description of the regulatory pathway, its requirements and the distinction between investigational treatment and marketing approval. FDA guidance.
Phagos (October 20, 2026). Introducing CAPHARD: A Model to Find the Viruses Most Likely to Infect Your Bacteria. Contemporary example of AI-based phage–host prediction. Phagos announcement.
The Phage Therapy (October 20, 2026). Phagos Launches CAPHARD, a 900-Million-Parameter AI Model to Predict Phage–Bacteria Interactions. Read the related article
.png){kind=logo}
Comments
Post a Comment