The Phage Therapy Ecosystem: From AI to Patient

Bringing a new treatment to patients typically involves years of research, clinical trials, and regulatory review before it becomes widely available. However, under exceptional circumstances, patients may receive an investigational treatment before this process is complete.

For patients facing serious conditions, standard treatments may have failed or may no longer be suitable. When satisfactory alternatives are unavailable and participation in a clinical trial is not possible, the FDA provides a pathway called expanded access², commonly known as compassionate use. This allows patients to receive investigational treatments that have not yet been approved for general use, outside of a clinical trial.

For an individual patient, the treating physician typically submits an Investigational New Drug (IND) application to the FDA, which reviews the proposed treatment and determines whether it may proceed. The central consideration is whether the potential benefit justifies the risks for the individual patient.

Compassionate use is not unique to phage therapy and has also been used for investigational cancer treatments³. While investigational treatments carry uncertainties, phage therapy has accumulated substantial clinical experience and generally favorable safety⁴ evidence⁵.

The authorization timeline depends on the urgency of the case, with emergency INDs (eINDs) sometimes authorized the same day. New non-emergency single-patient INDs generally involve a 30-day waiting period⁶, though the FDA can allow treatment sooner. Requirements also vary internationally. Canada has its Special Access Program, while European countries have their own national pathways. Radical Numerics works with partners in the US, Canada and Europe to navigate these requirements and help translate phage science into patient care.

Even with regulatory pathways such as compassionate use, phage therapy remains difficult for patients to access. As it is not yet routinely available through established clinical referral networks, treating physicians often need to seek out specialists or phage centers themselves.

Only a handful of phage therapy centers operate in the US, helping assess cases and connect treating physicians with laboratories and manufacturers. Some operate primarily as consultants, with the patient's own physician continuing to direct care. UC San Diego's IPATH, for example, follows this model⁷.

Much of this work also relies on research networks and personal connections. A clinician might contact a researcher with a phage collection, who in turn connects them with another laboratory that may have relevant candidates. Community networks such as Phage Directory⁸ have extended these searches across institutions and countries, but finding laboratories with the capacity to respond remains a challenge. A recent American Society for Microbiology initiative⁹ identified fragmented handoffs and limited capacity as major barriers, with phage matching and production of clinical-grade material often taking longer than regulatory paperwork.

Sustaining this infrastructure also presents a financial challenge. The FDA can permit recovery of certain direct product costs, but this provides a limited foundation for maintaining phage collections, staffing, and coordination across cases. Philanthropy and institutional support help fill the gap, with some programs providing phages free of charge and relying on donations to support their work.

Even when a suitable phage collection is available, finding a match for a patient's bacterial isolate can be challenging. Phages differ widely in their host range, or the bacteria they can infect, with many infecting only a subset of strains within a species (Figure 2A). This specificity enables targeted treatment but means that a phage active against one patient's isolate may have no activity against another^10,11.

Many phages are originally discovered in environments where their bacterial hosts live, including wastewater, surface water, and soil. Researchers test environmental samples against bacterial cultures, purify clonal phage populations, and propagate them to build stocks for characterization and inclusion in phage libraries^12,13.

These libraries help avoid starting from scratch for every patient. Academic and commercial phage banks, internal collections, and other laboratories provide existing candidates for testing. However, their usefulness depends on their coverage of relevant bacterial strains, not simply the number of phages they contain^13,14.

Personalized selection begins by culturing the infecting bacterium and screening phages against that isolate. When banked candidates fall short, targeted environmental searches may identify additional options. IPATH estimates that matching typically takes one to four weeks after bacterial isolates are shipped, with no guarantee of finding a match^13,14. However, finding a matching phage is only one step toward treatment. In a published analysis of IPATH's early cases¹⁵, the median time from requesting phage therapy to administration was approximately 171 days, ranging from 28 to 386 days. This reflects the additional time required for phage preparation, regulatory review, and clinical coordination.

Looking ahead, we hope Omnii-1's phage-host interaction predictions can help prioritize candidates when characterization data are limited, potentially reducing experimental screening effort (Figure 2B). Even when conventional screening and further searches yield no suitable match, they may identify phages worth engineering. Omnii-1 design and prioritization results¹

We begin by selecting a starting phage, or chassis, that has the potential to infect the target bacterial species but may not be active against the patient's particular strain. We select these chassis based on biological evidence and engineering feasibility, with Omnii-1's phage–host predictions providing additional guidance. For a potential cocktail, we may choose phages that target the bacterium through different mechanisms. If a bacterial change confers resistance to one phage, another may remain active. This provides a basis for designing cocktails that are less vulnerable to bacterial escape.

After selecting a chassis, we need both the genome sequence and the physical DNA itself. The genome may come from a collaborator, an existing phage library, or de novo chemical or enzymatic DNA synthesis. For some recent E. coli experiments, we sourced a purified T7 genome from a vendor for $50²⁰!

For each chassis, we identify the functions we want to modify and define regions of interest, the stretches of DNA we plan to redesign (Figure 3A). To change which strains a phage recognizes, for example, we might target sequences that encode the parts of a receptor-binding protein that contact the bacterium. Genome comparisons, structural analysis, and prior experiments help define the boundaries of each region. Where prior evidence is limited, we aim to use Omnii-1 to help identify and prioritize regions for investigation.

Once the regions of interest are defined, Omnii-1 generates, screens, and ranks phage variants in silico. Comparing predicted and measured activity helps evaluate how well the model prioritizes candidates (Figure 3B). The goal is to preserve phage function while improving activity against the target strains for effective downstream therapy. We combine the deliverables from Omnii-1 with quality checks to assemble experimental libraries that include top candidates, diverse alternatives, and controls. Omnii-1 design and prioritization results¹

Omnii-1’s designs then move into the lab, where high-throughput assays measure variant recovery and activity against the target isolate. Growth assays assess how selected phages suppress bacterial growth (Figure 3C). These results feed back into Omnii-1 to refine predictions and guide subsequent design rounds. As experimental data and computational capacity grow, we can improve AI-design and move more into screening in silico, reducing the lab iterations needed to improve and select candidates.

The experimental work starts with constructing variant libraries containing tens of thousands of designs generated by Omnii-1. Because the modified genes are relatively short, we can synthesize these designs efficiently using oligo pools, which allow thousands of distinct DNA sequences to be produced together in a single batch at a fraction of the cost of synthesizing each individually. The resulting sequences are then incorporated into the phage chassis using established molecular biology techniques, producing complete phage genomes, each carrying a distinct Omnii-1 design.

Once constructed, the engineered genomes must be rebooted into functional phages, a process that recovers infectious phage particles from their DNA. We have explored methods ranging from in vitro transcription and translation to electroporation into host cells, favoring those that maintain relatively even representation of library members for a given phage–host combination. Some designs fail to reboot, potentially because the resulting phages cannot assemble successfully.

We then screen the recovered phage libraries against bacterial host strains. Variants that proliferate during screening are identified through genome sequencing, allowing us to measure their relative enrichment. Because sequencing alone measures relative performance, we also use titration and plaque assays to assess infectious phage activity.

Promising candidates from pooled screening are then evaluated individually against relevant bacterial isolates. We measure how readily bacteria develop resistance to individual phages and test combinations to identify promising cocktails. Because phages can complement one another or interfere with each other's activity, the strongest individual candidates do not necessarily produce the strongest cocktail.

Phage cocktails also create opportunities for collaboration. Rather than relying exclusively on phages from a single organization, we can work with partners to combine naturally occurring phages, existing phage collections, and AI-designed variants into complementary cocktails. By sharing experimental data and evaluating how different phages perform together, we aim to help design combinations that expand bacterial coverage and reduce the likelihood of resistance.

This work involves a network of specialized collaborators. We work with DNA synthesis providers such as Twist and Ansa, and experimental biology partners such as Flock Bio and Invitris to help construct, reboot, and evaluate phages. Other collaborators contribute phage collections, bacterial isolates, cocktail development, and connections to clinical teams. Each plays a distinct role in translating computational designs into physical phages that can be evaluated for potential clinical use.

Even after promising phages and cocktails are identified experimentally, substantial work remains before they can be administered to a patient. The selected phages must be produced and purified in sufficient quantities, then tested for clinical suitability. Key assessments include sterility and endotoxin testing, the latter measuring a bacterial component that can cause harmful inflammatory responses. Genome sequencing helps verify the intended design and screen for concerning genetic features, while activity and purity testing provide further evidence of product quality. These are typically outsourced to vendors that specialize in these standardized protocols (e.g. USP 71 for sterility testing).

Once the phage preparation meets the required quality specifications, it must reach the treating hospital with appropriate documentation, storage and handling instructions, and sufficient material for treatment. Shipping and pharmacy preparation must align with the clinical schedule before administration.

Today, personalized phage therapy depends on a fragmented network of researchers, clinicians, manufacturers, and specialized laboratories. Each patient case can require extensive coordination, from identifying suitable phages to experimental validation and clinical preparation.

At Radical Numerics, our goal is to make this process more systematic by advancing AI-driven capabilities, from phage–host interaction prediction to targeted phage generation, alongside scalable experimental validation, and working closely with partners across manufacturing and clinical care. We believe that by bringing these capabilities and expertise together, we can all help establish a more reliable and repeatable pathway for personalized phage therapy, expanding access for patients facing difficult infections.

We are excited to build on the progress of the broader phage therapy community and welcome collaborators who share our goal of making these treatments more accessible to patients.