As antimicrobial resistance (AMR) narrows treatment options for some bacterial infections, bacteriophages are attracting renewed therapeutic interest because they can target bacteria with extraordinary specificity. That same specificity creates one of the central operational challenges of phage therapy: the therapeutic value of a phage can depend on finding the right match for the bacterial strain present in an individual patient. Personalized programs already select phages against patients’ bacterial isolates, and regulators have identified bacteriophage therapy as an example of the broader move toward individualized therapeutics and precision medicine.1,2
An August 2026 study in Scientific Reports shows why that matching problem may be more complicated than simply determining whether a pathogen is susceptible.3 Researchers tested the lytic phage HMGUpa1 against five clinical isolates of Pseudomonas aeruginosa. The phage could infect all five strains, but its killing efficiency varied sharply, from approximately 90% against one strain to approximately 60%, 30%, 20%, and close to zero against the others. The strains themselves were highly similar at the genomic level, with average nucleotide identity above 99% and average amino acid identity above 98%. Differences in adsorption, latency, and other conventional infection parameters did not consistently explain the variation. Instead, differential responses were associated with a wider set of host characteristics, including transcriptional changes involving secretion systems, RNA repair, metabolism, sulfate transport, envelope stress, and other pathways.
The study does not establish a molecular signature that can predict whether a phage will work clinically. It does, however, expose the limitation of treating phage susceptibility as a simple binary property. If the same phage can infect several closely related strains but produce dramatically different levels of bacterial killing, the more consequential question is not merely whether the pathogen is susceptible, but how good the match actually is.
From Host Recognition to Bacterial Killing
Phage activity depends on a sequence of biological events. A phage must recognize an appropriate bacterial host, successfully infect it, replicate productively, and ultimately kill or suppress enough bacteria to meaningfully affect the population. Success at an early stage does not guarantee success downstream.
Current computational approaches make this distinction especially visible. One strain-level prediction system developed for Klebsiella species uses bacterial and phage genomic information to rank likely phage–host interactions, focusing in part on receptor-binding proteins and their bacterial targets.4 The predictions identify promising candidates for experimental validation rather than replace laboratory testing. A visible interaction in a spot assay can indicate an initial phage–bacterium interaction without necessarily demonstrating productive replication.
Host range is also shaped by more than receptor recognition. Bacterial defense systems, environmental conditions, interactions among phages and hosts, coevolution, and phenotypic variability can affect whether infection succeeds, and host range can change rather than remaining a fixed property of a phage. Genomic tools are increasingly being developed to predict these relationships at the strain level.5
The P. aeruginosa findings add another layer. Even after the same phage infected all five strains, antibacterial performance remained highly variable. Host-range testing can identify potential candidates without fully describing how effectively those candidates will act.3
The Target May Be More Than One Strain
The most immediate implication is that phage matching may need to occur at the strain level. A phage described broadly as active against P. aeruginosa may perform very differently against individual P. aeruginosa isolates, even when those isolates are genetically similar. Host-specific molecular and physiological traits may help determine the quality of that interaction.
But even strain-level matching may not capture the entire clinical problem.
A 2025 study from the Journal of Applied Microbiology following an methicillin-resistant Staphylococcus aureus (MRSA) case compared phage responses using individual bacterial isolates with responses generated from mixtures of isolates recovered from the same patient.6 Mixed populations were more likely to persist under phage exposure than individual isolates, and, at one infection site, experiments using those mixtures better reproduced the clinical behavior observed during treatment. The relationship also differed according to anatomical site.
One case cannot establish how often within-patient heterogeneity will alter treatment response, but it raises an important question about sampling. The issue is not only whether testing reaches the strain level, but whether the strain selected for testing adequately represents the bacterial population being treated. If it does not, even a strain-level phagogram, the susceptibility profile showing how a bacterial isolate responds to candidate phages, may provide an incomplete picture of expected therapeutic activity.
The Phagogram May Need to Become More Informative
Phage susceptibility testing (PST), sometimes described through the use of a phagogram, is already central to personalized phage selection. However, commonly used plaque-based and growth-kinetic methods remain labor-intensive and time-consuming, and no easy standardized reference method has been established.7
The larger question is what a future phagogram should actually measure. A binary susceptible/resistant designation may not capture how strongly, reliably, or durably a particular phage suppresses a particular bacterial target.
Current reproducibility challenges reinforce that point. In the absence of an established reference standard for PST, one important requirement for any future phagogram is that the same phage–bacterium pairing produce reasonably consistent results when tested in different laboratories. A multicenter study examined that problem directly by testing four phages and a four-phage cocktail against 145 P. aeruginosa respiratory isolates, including 46 from people with cystic fibrosis. Two laboratories used a liquid-based assay that had been standardized between the sites, while a third used a double-layer plaque assay. Even under those conditions, agreement between the two laboratories using the same liquid method was only fair, with concordance correlation coefficients below 0.6 for every phage tested, and agreement was poorer still when liquid-assay results were compared with the plaque method.8
Clinical implementation may also require a broader testing infrastructure, including standardized PST methods and controls, automation for screening large phage libraries, and repeat testing when susceptibility changes during treatment.9
Computational tools could help make that richer testing strategy practical by narrowing the experimental search space. Strain-level prediction models already demonstrate the potential to use genomic information to rank phage candidates before laboratory screening, reducing the number of combinations that need to be tested experimentally. They do not yet eliminate the need for phenotypic validation.4
The new P. aeruginosa research suggests why that validation remains important. Differences in killing were associated with multiple host characteristics rather than one simple determinant that could be read directly from a genome. The study links efficacy with differences in bacterial transcription, physiology, and infection behavior, but it does not establish a validated method for predicting killing from those features.3
A future precision-matching system could therefore integrate bacterial characterization, computational candidate ranking, phenotypic susceptibility testing, and quantitative measures of killing into a more informative ranking of available phages. The objective would be to identify which candidates are most likely to produce meaningful antibacterial activity against the bacterial population actually being treated.
A Good Match May Not Stay a Good Match
Even a well-characterized pre-treatment match may not remain static once therapy begins.
In a retrospective series of 100 personalized phage-therapy cases, emergence of phage-insensitive bacteria was documented in seven of the 16 patients in whom resistance was evaluated. Some resistant isolates also showed antibiotic re-sensitization or reductions in virulence-related traits. Because the study was observational and uncontrolled, those findings do not establish clinical efficacy, but they demonstrate that the phage–bacterium relationship can change during treatment.1
Time therefore becomes another dimension of precision. If therapy changes the composition or characteristics of the bacterial population, susceptibility determined before treatment may become outdated. Clinical laboratories may need the ability to repeat testing when breakthrough growth or resistance appears rather than treating the initial phagogram as a permanent characterization of the infection.9
Phage selection could eventually incorporate the consequences of that evolutionary pressure as well. In a small 2025 personalized therapy study published in Nature Medicine involving nine adults with cystic fibrosis and multidrug- or pan-drug-resistant P. aeruginosa, phages were selected partly according to predicted evolutionary trade-offs.10 The strategy sought phages for which development of resistance could impose changes that reduced bacterial virulence or antibiotic resistance, and the study found evidence of such trade-offs while emphasizing the need for evaluation in larger clinical trials.
This broadens what may eventually count as a useful match. Immediate killing may not be the only relevant property. The bacterial response selected by the phage could also become part of therapeutic decision-making.
Precision Does Not Necessarily Mean Bespoke Manufacturing
Greater biological precision creates an apparent scalability problem. If phage selection has to occur at the strain level, account for population heterogeneity, and sometimes change during treatment, it is easy to assume that every patient would require a newly manufactured therapeutic product.
Existing models suggest a more flexible path.
Clinical literature distinguishes among non-customized phage therapy, customized selection of phages from existing collections, and customized production in which phages are additionally propagated using the patient’s bacterial isolate. Customization can add time to treatment workflows, making the distinction between selecting an existing product and producing a new one practically important.10
Precision can therefore reside in the selection process without requiring every dose to be manufactured from scratch.
A deep, well-characterized phage inventory could provide multiple candidates for a given clinical isolate. Rapid bacterial characterization and computational ranking could narrow that library to the most promising options, followed by targeted laboratory testing to confirm activity. Existing phages or defined cocktails could then be deployed when an appropriate match is available, while isolation, adaptation, or individualized production could remain options when existing inventories do not provide a suitable candidate.
Elements of this model already exist in practice. A Belgian consortium involving Queen Astrid Military Hospital, KU Leuven, and Sciensano retrospectively analyzed its first 100 consecutive personalized phage-therapy cases, spanning difficult-to-treat infections managed at 35 hospitals in 12 countries between 2008 and 2022. Most involved salvage therapy after standard antibiotic treatment had failed. Rather than relying on a fixed product, the program selected treatments from 26 individual phages and six defined cocktails according to activity against the causative bacterial isolate, with some phages pre-adapted to improve activity against the intended target. Across the first 100 patients, 13 of the 26 individual phages used had undergone such adaptation.1
The model also required a manufacturing and quality infrastructure capable of converting those selections into patient-specific medicines. Queen Astrid Military Hospital produced the phage active pharmaceutical ingredients (APIs), while Sciensano evaluated their quality and safety according to Belgium’s phage API monograph, including genomic characterization, phage identity and purity, lytic activity, bioburden, pH, and bacterial endotoxin testing. More than 40 individual phage API batches were quality controlled for treatment of the first 100 patients. The experience therefore provides a useful early example of what personalized phage therapy may require operationally: not simply a sufficiently large phage library, but a connected system for susceptibility testing, phage selection and adaptation, manufacturing, quality control, and clinical deployment.
Other preparation models similarly aim to reduce the time between identifying an active phage and treating the patient. Yale’s Center for Phage Biology & Therapy has developed a modular pipeline for single-patient emergency investigational use against P. aeruginosa, with characterized cell banks and phage stocks, defined production and quality-testing steps, and the longer-term goal of maintaining larger collections of fully characterized, ready-to-use phages that can be rapidly matched to clinical isolates.11 Regulatory proposals based on Belgium’s magistral model take a complementary approach: pharmacies can prepare individualized phage combinations for named patients from qualified phage stocks, while industrially manufactured fixed cocktails could still serve indications where standardized or immediately available treatment is preferable.12
Regulators are already considering bacteriophage therapy within the broader category of individualized therapeutics and precision medicine.2 Personalized phage cocktails targeted to a patient’s bacterial strain illustrate the opportunities and challenges of individualized products, including the need for robust manufacturing and reliable assurance of product quality.
The manufacturing challenge is closely tied to the diagnostic one. The usefulness of a large phage bank depends on how rapidly and accurately clinicians can identify the best available match, while increasingly precise matching becomes clinically useful only if appropriate therapeutic material can be accessed quickly enough to act on the result.
From Susceptible to Predictable
The new strain-level study demonstrates why better prediction may eventually matter. Closely related bacterial strains exposed to the same phage can produce substantially different killing outcomes, and those differences appear to reflect a complex set of host characteristics.
Precision phage therapy may therefore require several successive questions. Can this phage productively infect this bacterial target? How efficiently does it suppress this particular isolate? Does the isolate adequately represent the infection? Does that relationship remain favorable as treatment changes the bacterial population?
Answering those questions quickly will require progress across diagnostics, computational prediction, phage libraries, manufacturing, and clinical laboratory workflows. The resulting model may be less about creating a unique phage for every patient than about making increasingly informed choices among well-characterized therapeutic options and knowing when those choices need to change.
“Phage susceptible” may ultimately describe the beginning of that process rather than its endpoint. The more demanding goal is to move from susceptible, to matched, to monitored, and, eventually, perhaps, predictable.
