Phage Therapy: Why It Hasn’t Gone Mainstream Yet

Phage therapy, treating bacterial infections with the viruses that prey on bacteria, is a century old and still has no naturally occurring phage product registered with either the US Food and Drug Administration or the European Medicines Agency, a point made in a 2025 review of the regulatory landscape (Morales and Hyman 2025). The gap is odd on its face. Drug-resistant infections keep rising, individual rescue cases make the news, and phages are abundant, specific and cheap to isolate. So what is actually holding the field back, and what would have to change for it to become ordinary medicine?

Treat that as an unsettled research question rather than a story with a known ending. Several explanations compete, and they imply different fixes: a biology problem, a trial-design problem, a manufacturing problem, or a legal-category problem. People outside the field ask the same thing in blunter terms. A 2016 thread on Physics Forums opens with a reader asking why phage treatments are not permitted in humans or taken seriously as an antibiotic alternative (Physics Forums 2016). A decade on, the honest answer is still that it depends which bottleneck you think binds.

Why the question has not been resolved

Two bodies of evidence point in opposite directions. Compassionate-use programmes report large fractions of desperately ill patients improving once phages are added to their regimen. The small number of randomised trials has been less encouraging. Each camp can discount the other: advocates argue the trials used the wrong product at the wrong dose, and critics reply that uncontrolled series flatter any intervention, because patients on concurrent antibiotics often recover anyway.

A definitional problem compounds this. The phrase covers at least four distinct things: fixed off-the-shelf cocktails, phages matched to a single patient’s isolate, genetically engineered phages, and purified phage enzymes such as lysins. Evidence about one format transfers poorly to another. Much of the apparent disagreement in the literature is really two groups describing different products under one name.

What the best-known randomised trial actually showed

PhagoBurn remains the most cited controlled test: a randomised, double-blind phase 1/2 trial across nine burn centres in France and Belgium (Jault et al. 2019; registry record NCT02116010). Patients whose burn wounds were infected with Pseudomonas aeruginosa got either a twelve-phage cocktail, PP1131, or the standard 1% silver sulfadiazine cream. Twenty-seven patients were randomised; twenty-five entered the efficacy analysis.

The outcome went the wrong way for the viruses. Median time to sustained reduction in bacterial burden was 144 hours with phages against 47 hours with cream (hazard ratio 0.29, 95% CI 0.10 to 0.79). Recruitment stopped on 2 January 2017 for insufficient efficacy. Safety looked fine: adverse events in 23% of phage patients against 54% on standard care, with one unrelated death in each arm.

Then comes the detail that makes the trial hard to read as a verdict on the biology. Titre fell during and after manufacturing, so patients received roughly 10² plaque-forming units per millilitre per dose rather than the intended 10⁶ — a shortfall of four orders of magnitude. Isolates from patients who failed were resistant to low phage doses. The authors stated their conclusion narrowly:

“At very low concentrations, PP1131 decreased bacterial burden in burn wounds at a slower pace than standard of care.”

Jault et al., 2019

Read strictly, PhagoBurn tested a badly underdosed batch. It is strong evidence that phage products are fragile to make and weak evidence about whether a correctly dosed cocktail works. That ambiguity, in one trial of 27 people, is roughly the state of the controlled record for topical phage treatment.

What a hundred personalized cases suggest

The counterweight is observational. A Belgian consortium published outcomes for its first 100 consecutive personalized cases, treated between 2008 and 2022 for patients in twelve countries with difficult-to-treat infections (Pirnay et al. 2024, Nature Microbiology). Clinical improvement was reported in 77.2% of infections and eradication of the target organism in 61.3%. Fifteen adverse events were recorded, seven of them non-serious reactions suspected of being linked to the treatment.

Three numbers in that paper matter more than the headline success rate. Eradication was about 70% less probable when no concomitant antibiotic was given (odds ratio 0.3, 95% CI 0.127 to 0.749). In vivo selection of phage resistance was documented in 43.8% of evaluated patients, seven of sixteen. Phage–antibiotic synergy appeared in vitro in nine of ten isolates tested, and immune neutralisation of the phages turned up in five of thirteen patients screened.

So the most informative large series does not describe phages as a replacement for antibiotics. It describes them as a partner drug, prone to resistance within a single course, sometimes cleared by the patient’s own antibodies. The design was uncontrolled, and the authors say so; selection of which patients got treated, and publication of which cases, both push success rates upward.

The regulatory fit problem

Both the US and EU definitions of a medicinal product hinge on intended use: something offered to diagnose, prevent, cure or mitigate disease is a drug, and must be made to current good manufacturing practice. A fixed cocktail can meet that standard. A product assembled this week from a library, against this patient’s isolate, cannot easily, because each batch is in effect a new drug. Morales and Hyman describe stakeholders viewing those quality standards as costly for

“a product whose composition could be, at times, specific to an individual patient”

Morales and Hyman, 2025

Belgium’s workaround is the magistral preparation: a framework in which a pharmacist compounds a tailored phage product against a monograph, with quality testing performed by the national public health institute. Morales and Hyman report that users of the route still face a patient backlog, mostly from logistics rather than law. Elsewhere, clinicians reach patients through expanded access in the US, the Special Access Scheme in Australia, or European compassionate use, all of which impose lighter production requirements and none of which generate the comparative data a licence needs.

Hendrik Brüssow, surveying the obstacles in Viruses, put the commercial consequence plainly:

“Patent and regulatory issues still cause some hesitation in the private sector.”

Brüssow, 2019

His list of hurdles is broader than paperwork. Phages replicate, vary and evolve, so defining a product’s composition is not trivial. Their pharmacokinetics do not fit standard dosing frameworks. Acute-infection trials are awkward because treatment must start before the microbiology comes back, which for a pathogen-specific agent is a real design bind. This is the same category trouble that shows up whenever a therapy is made per patient; the pathway questions raised by bespoke gene editing rhyme closely, as does the batch-definition problem in biosimilar manufacturing.

Four competing explanations, steelmanned

The hypotheses below are not mutually exclusive, but they rank the bottleneck differently, and that ranking determines where money should go.

1. It is a biology problem

On this view phages are intrinsically awkward therapeutics: narrow host range, resistance arising within days, antibody neutralisation, poor penetration of biofilms and mucus, and dependence on bacterial density for amplification. Evidence for: the resistance and neutralisation rates in the Belgian series. Evidence against: those same patients mostly improved, and resistance is hardly disqualifying — antibiotics select resistance too, and bacterial defence systems are themselves a moving target, as work on how phage proteases trigger CBASS immunity illustrates.

2. It is a manufacturing problem

Here the binding constraint is making a stable, endotoxin-free, correctly titred product and keeping it potent to the bedside. The four-log titre loss in PhagoBurn is the strongest single datum in favour, and it is hard to dismiss: the trial that most shaped opinion failed at the pharmacy, not the patient. Against it, Georgian and Belgian centres have supplied usable preparations for years, so the problem looks solvable rather than fundamental.

3. It is a trial-design problem

Perhaps the field keeps running the wrong experiment. A pathogen-specific agent tested in a syndrome-defined population dilutes any effect, since some enrolled patients carry organisms the cocktail cannot touch. Phage-susceptibility screening before enrolment, which Jault and colleagues recommended, fixes that at the cost of slower recruitment. The counterargument is uncomfortable: a therapy that cannot be tested in a conventional trial also cannot be licensed by conventional means, and that is a fact about the therapy, not only about the trial.

4. It is an economics-and-law problem

The last hypothesis says the science is adequate and the incentives are not. Natural phages are difficult to patent attractively, personalized preparations do not scale, antibiotic-adjacent markets have repeatedly bankrupted their developers, and no regulator has defined a dedicated category. Brüssow’s judgement is that current evidence is too thin to attract heavy private money, which implies public funding should lead. Against: Belgium shows a legal route can be built without new statute, and the backlog there is operational.

A small simulation: why cocktails delay resistance

Resistance recurs in both the trial and the case series, and it is the one piece of the argument that can be probed without a laboratory. The model below is a small simulation built for this article, not lab work and not a clinical prediction. It tracks a bacterial population in which every cell has a genotype recording which phages in the cocktail it resists, with a 5% growth penalty per resistance allele, logistic growth to a carrying capacity of 10⁹ per millilitre, mass-action infection, a burst size of 50, phage decay at 0.2 per hour, and resistance mutations arising at 10^-6 per cell division per phage. Each scenario is 200 stochastic runs over 72 simulated hours, seeded for reproducibility. A run counts as treatment failure when a genotype resistant to every phage in the cocktail reaches 10⁶ cells per millilitre.

The point of interest is combinatorial. Resistance to one phage needs one mutation; resistance to a three-phage cocktail whose members use independent receptors needs three in the same lineage, and the per-lineage probability of that is the product of three small numbers. The model simply makes that arithmetic concrete under finite population sizes and finite time.

import numpy as np, itertools
r, K, a, burst, dp = 1.0, 1e9, 1e-8, 50, 0.2   # /h, cells/mL, mL/h, phage/cell, /h
mu, cost, B0, T, dt = 1e-6, 0.05, 1e7, 72.0, 0.01

def run(k, dose_each, shared=False, reps=200, seed=1):
    rng = np.random.default_rng(seed)
    G = list(itertools.product([0, 1], repeat=k))   # 1 = resistant to phage j
    R = np.array(G, float)
    grow = r * (1 - cost) ** R.sum(1)               # cost per resistance allele
    B = np.zeros((reps, len(G))); B[:, 0] = B0      # all cells start susceptible
    P = np.full((reps, k), float(dose_each))
    t_res = np.full(reps, np.nan)
    for step in range(int(T / dt)):
        N = B.sum(1, keepdims=True)
        births = np.maximum(grow * B * (1 - N / K), 0) * dt
        inf = a * B[:, :, None] * (1 - R)[None] * P[:, None, :] * dt
        killed = np.minimum(inf.sum(2), B)          # cap kills at cells present
        scale = np.where(inf.sum(2) > 0, killed / np.maximum(inf.sum(2), 1e-300), 0)
        inf *= scale[:, :, None]
        B += births - killed
        P += burst * inf.sum(1) - dp * P * dt       # phage amplify on infection
        for gi, g in enumerate(G):                  # mutation to resistance
            for j in range(k):
                if g[j] == 0:
                    tgt = tuple(1 for _ in g) if shared else g[:j] + (1,) + g[j+1:]
                    n = rng.poisson(mu * births[:, gi])
                    B[:, gi] -= n; B[:, G.index(tgt)] += n
                    if shared: break
        B[B < 1] = 0                                # extinction floor
        hit = np.isnan(t_res) & (B[:, -1] >= 1e6)   # fully resistant genotype wins
        t_res[hit] = step * dt
    return t_res, B.sum(1)

for label, k, dose, sh in [("1 phage", 1, 1e8, False),
                           ("2-phage cocktail", 2, 1e8, False),
                           ("3-phage cocktail", 3, 1e8, False),
                           ("3 phages, same total dose", 3, 1e8 / 3, False),
                           ("3 phages, shared receptor", 3, 1e8, True)]:
    t, N = run(k, dose, sh)
    f = np.mean(~np.isnan(t))
    print(label, round(100 * f, 1), "%", None if f == 0 else round(float(np.nanmedian(t)), 1))

With a single phage at 10⁸ per millilitre, a resistant population took over in 197 of 200 runs, at a median of 13.3 simulated hours. Adding a second independent phage dropped that to zero runs in 200, as did three. Splitting the same total dose three ways also gave zero failures, which is the interesting result: the benefit came from the number of independent attack routes rather than from more virions. When all three phages were made to depend on the same receptor, so that one mutation conferred resistance to the whole cocktail, failure returned in 89% of runs. Raising the mutation rate tenfold to 10^-5 and the starting population to 10⁹ made single-phage failure universal (100% of runs) and still produced no cocktail failures.

It shows that under these assumptions, independence of resistance mechanisms is the property that buys time, and that cocktail size is a poor proxy for it. A twelve-phage cocktail whose members all bind one surface structure is, in this model, roughly a one-phage cocktail. That is a sharper claim than the familiar advice to use several phages, and it is checkable in vitro by receptor mapping rather than by counting components. It is consistent with the older argument that cocktails can forestall resistance (Keen 2012), while adding the condition that makes it fail.

What it does not show is larger. The model has no immune system, no biofilm, no spatial structure, no lysogeny, no plasmid-borne or CRISPR-based defence, and no horizontal transfer of resistance between lineages. It takes mutation rates and fitness costs as fixed inputs rather than measuring them. Real resistance is often not a clean receptor knockout but a partial change with pleiotropic costs. So the result is a statement about the combinatorics of independent targets, nothing more, and it cannot be used to pick a dose or predict a patient’s course. Why cocktails and drug combinations work at all rests on the same logic that governs mutation, selection and drift more generally.

My reading of the evidence

What follows is the author’s assessment, not a finding. I think the manufacturing-and-category explanation is doing most of the work, and the pure biology explanation is the weakest of the four. The reasoning is simple: the trial that most damaged the field’s reputation delivered about one ten-thousandth of its intended dose, and a programme that controls its own compounding reports usable outcomes in the majority of a hundred consecutive patients. Those two facts sit badly with the claim that phages cannot work in bodies.

Second opinion, held less firmly: the field’s framing of phages as antibiotic replacements has cost it years. The Belgian odds ratio points the other way, towards phages as adjuncts given alongside antibiotics, and the in vitro synergy rate supports that. A combination indication is harder to design and easier to win, because the comparator is a real standard of care rather than nothing.

Third, the legal bottleneck is narrower than commonly claimed. Belgium built a route under existing pharmacy law. The remaining friction there is reported as logistical. That suggests the scarce resource is not statute but capacity: qualified compounding facilities, validated potency assays, phage libraries with receptor annotation, and reimbursement codes.

What would settle it

Specific, falsifiable things that would move the question, with predictions attached:

1. A randomised trial that enrols only patients whose isolate is phage-susceptible on a pre-treatment phagogram, with release titre verified at the bedside rather than at manufacture. Prediction: such a trial shows a significant effect on microbiological endpoints. If a properly titred, susceptibility-screened trial still fails, the biology hypothesis gains a great deal.
2. A head-to-head comparison of cocktails matched for component count but differing in receptor independence. Prediction, from the model above: time to resistance tracks the number of distinct receptors, not the number of phages.
3. Routine reporting of anti-phage antibody titres and in-patient resistance in every treated case, so that the 43.8% and 38.5% figures from the Belgian series can be confirmed or revised in a larger denominator.
4. A published stability dossier explaining the PhagoBurn titre loss. If a standard production route cannot hold titre, that is a finding worth as much as a trial.
5. Registry data from a magistral programme with denominators, including patients who were considered and not treated. This is the only realistic way to estimate selection bias in case series.

None of these needs a new scientific idea; each needs funding for unglamorous work on potency, receptors and record-keeping. This article is general information about regulation, not legal advice.

Common questions

Is phage therapy legal anywhere?

Yes, by several routes, none of which is a marketing licence. Belgium permits tailored phage preparations compounded by a pharmacist under a national monograph, with quality control by the public health institute. The US, Australia and the EU all have access pathways for unapproved medicines in patients out of options. Morales and Hyman report that no naturally occurring phage product is registered with the FDA or EMA.

Did PhagoBurn prove phage therapy fails?

No, and the trial report does not claim that. Patients received roughly 10² plaque-forming units per millilitre instead of the planned 10⁶, because titre dropped after manufacturing. The authors concluded that the cocktail worked more slowly than standard care at very low concentrations and called for studies at higher concentrations in larger samples.

Do bacteria become resistant to phages?

Routinely. The Belgian series documented in vivo selection of phage resistance in 43.8% of evaluated patients. Resistance usually involves losing or altering the surface structure a phage binds, which often carries a fitness cost. Cocktails targeting different receptors are the standard countermeasure, and the simulation above suggests independence of those receptors matters more than the number of phages.

Are phages a replacement for antibiotics?

The strongest observational data argue against that framing. In the hundred-case series, eradication was about 70% less probable when no concomitant antibiotic was used, and in vitro synergy between phage and antibiotic appeared in nine of ten isolates tested. On current evidence the plausible role is as an adjunct in infections that conventional drugs cannot clear.

Why does personalization complicate approval?

Drug law in both the US and EU defines a medicinal product by intended use and then requires each batch to be made and characterised to a fixed specification. A preparation matched to one patient’s isolate is, in regulatory terms, a new product each time, which is why compounding and access pathways have carried most of the clinical activity so far.

References

1. Jault, P., Leclerc, T., Jennes, S., Pirnay, J.-P., Que, Y.-A., Resch, G., et al. (2019). Efficacy and tolerability of a cocktail of bacteriophages to treat burn wounds infected by Pseudomonas aeruginosa (PhagoBurn): a randomised, controlled, double-blind phase 1/2 trial. The Lancet Infectious Diseases, 19(1), 35–45. doi:10.1016/S1473-3099(18)30482-1 (record: boris.unibe.ch/120908)
2. ClinicalTrials.gov. Evaluation of phage therapy for the treatment of Escherichia coli and Pseudomonas aeruginosa wound infections in burned patients, NCT02116010. clinicaltrials.gov/study/NCT02116010
3. Pirnay, J.-P., Djebara, S., Steurs, G., et al. (2024). Personalized bacteriophage therapy outcomes for 100 consecutive cases: a multicentre, multinational, retrospective observational study. Nature Microbiology, 9(6), 1434–1453. doi:10.1038/s41564-024-01705-x
4. Brüssow, H. (2019). Hurdles for phage therapy to become a reality—an editorial comment. Viruses, 11(6), 557. PMC6631134
5. Morales, S., & Hyman, P. (2025). Quis custodiet? Are regulations slowing phage therapy? Drugs and Drug Candidates, 4(1), 1. mdpi.com/2813-2998/4/1/1
6. Keen, E. C. (2012). Phage therapy: concept to cure. Frontiers in Microbiology, 3, 238. doi:10.3389/fmicb.2012.00238
7. Physics Forums (2016). Bacteriophage therapy: alternatives to antibiotics (discussion thread, 23 July 2016). physicsforums.com