Towards modeling phage therapy Authors: Rob J. de Boer, Robert Schooley, Alan S. Perelson, Roland Regoes, James Faeder, Roland Regoes, James Faeder, Roland Regoes, James Faeder, Roland Regoes, James Faeder, Roland Regoes, James Faeder Published: 2026 ABSTRACT Patients infected with life-threatening multi-drug resistant (MDR) bacteria have been treated with cocktails of bacteriophages. This is a complicated form of personalized medicine as the phages given to a patient have to be selected beforehand on the basis of their lytic capacity of the infecting bacteria. Because bacteria rapidly become resistant, the evolution of resistance to a diverse cocktail of phages is a complicated dynamical process, during which competing bacterial strains replace one another by accumulating several resistance mechanisms, each of which may involve a fitness cost. As a consequence, it is typically not known why a particular phage therapy succeeded or failed, and how one can optimize the composition of the cocktails to maximize the rate of success. To improve upon this, we extend an existing in vivo-calibrated mouse model into a novel mathematical model for the human situation, and include multiple phages infecting multiple bacterial strains, differing in their resistance to each of the phages. We adjust several parameter estimates of the bacterial model to the human situation, and use the model to describe a successful case of phage therapy involving several cocktails, each containing several phages. In the model, treatment success crucially depended on pretreatment resistance levels, and on the diversity and the timing of the cocktails. Once an appropriate cocktail is found, it is less important to further optimize the infection rates of the phages. Resistant bacterial strains expand rapidly when sensitive strains decline, and the higher the infectivity of the phages, the faster resistant strains expand. Because resistance evolves rapidly, it is best to provide a diverse set of phages right from the start of therapy, i.e., to hit hard and early, and create a high genetic barrier to bacterial resistance. 1 Introduction Due to the evolution of bacterial resistance to multiple types of antibiotics, viruses that can infect and kill bacteria (called bacteriophages), are receiving renewed interest as a treatment option for bacterial infections. Recent clinical successes with patients recovering from rampant bacterial infections due to personalized phage therapy, have sparked interest into the design, efficacy, and safety of cocktails of phages that can be given to patients [1,2]. Phage therapy has been used as ‘living antibiotics’ in the Soviet Union over extended periods of time [3], but these recent successes were ‘compassionate use’ cases of life threatening bacterial infections that were treated with personalized cocktails of phages hastily selected based on their capacity to lyse the bacteria causing the infection [4–7]. Unfortunately, the very limited host range of most bacteriophages calls for such a personalized approach, and is a major challenge for generalized treatments and clinical trials [1]. Phage therapy may work for several reasons: (1) phages may simply infect and eliminate all bacteria (which is unlikely due to the rapid evolution of resistance [8,9]), (2) phages may reduce bacterial concentrations to levels where the host immune system regains control [10], (3) by evolving resistance to phages bacteria may regain sensitivity to particular antibiotics [2,4,5,11], (4) by modifying their capsule to prevent infection [7] bacteria may become more sensitive to killing by neutrophils, and (5) bacteria that evolved phage resistance may have such a crippled fitness that their replication rate drops below the rate at which the host immune system eliminates them. It seems virtually impossible to distinguish between these mechanisms on the basis of observational clinical data alone [1,6]. Because we expect many more people with multiple drug resistant (MDR) infections to be treated with phage therapy in the near future [2], we here develop mathematical models implementing these mechanisms, with the aim to employ these models for better interpretation of future data. Such models may help one optimize the design of therapy, and by fitting variations of these models to the data, one may become able to identify the mechanism(s) underlying the control of rampant bacterial infections by phage therapy in individual cases. The infection of bacteria by phages, and the rapid evolution of resistance by bacteria, has been studied extensively in vitro, and it is known that these systems tend to a approach an attractor where susceptible and resistant bacteria co-exist with the phages [8,9]. It was therefore surprising that phage therapy can be successful in vivo, and by elegantly combining modeling with experiments in mice, Roach et al. showed that the innate immune response, in this case mostly neutrophils, is largely responsible for the in vivo success of phage therapy in their experiments [10,9]. Apparently, neutrophils are usually incapable of controlling rampant bacterial infections, and a major effect of phage therapy is to reduce bacterial populations to a level where the killing of bacteria by neutrophils suffices to control bacterial growth. This mechanism requires the existence of a threshold in the bacterial density above which neutrophils can no longer control bacterial growth. Such a threshold is readily brought about in a model where at high bacterial densities (1) the neutrophils approach a maximum density, and (2) the number of bacteria killed per neutrophil per unit of time approaches a maximum, e.g., due to the fact that killing takes time [10,12,13]. Neutrophil densities rapidly increase during inflammation [14]. Because they are released more rapidly from the bone marrow, they can approach a new (higher) steady state level on a time scale of less than a day [15,16]. Thus, there should be two thresholds: (1) the maximum bacterial density that can be controlled by the normal density of neutrophils, e.g., small bacterial invasions that are controlled on a daily basis without triggering additional inflammation, and (2) the much higher threshold above which infections can grow uncontrolled despite the massive killing by neutrophils at their maximal density. In between an innate immune response is required to control the infection before it breaches the upper threshold. After developing a novel mathematical model having both thresholds, and allowing for several bacterial strains differing in their sensitivity to various phages and antibiotics, we illustrate its usefulness by describing a case of successful phage therapy of a patient suffering from a rampant infection with multi-drug resistant (MDR) Acinetobacter baumannii bacteria [3,4,17]. We present a few scenarios suggesting why this particular phage therapy might have been so successful. This patient was treated with several phage cocktails that were given sequentially, leading to rapid sequential development of bacterial resistance [7]. Most importantly, we find that the level of resistance increases rapidly when sensitive bacteria decline due to the (unsuccessful) phages in the first cocktail, and that this may markedly reduce the efficacy of subsequent cocktails. Thus, similar to lessons learned from treating rapidly evolving viruses like HIV, we find that it is important to ‘hit hard and early’ [18] by providing all phages at the same time, and to increase the ‘genetic barrier’ [19] by providing as many phages as possible. 2 Results The killing of bacteria by neutrophils has been studied widely in vitro [20–23] and in vivo [24,25]. Importantly, these data reveal that there are two types of thresholds, one for the neutrophils and one for the bacteria. The control of bacterial growth requires a minimum number of neutrophils, which has been called the critical neutrophil concentration (CNC) [20–22]. The in vitro data in these papers have been modeled successfully with exponential growth of bacteria, B, and mass-action killing by neutrophils, N, i.e., dBdt=rB−kNB, revealing that the neutrophil density at which dBdt=0, i.e., N^=r/k, defines a CNC that is independent of the bacterial density [20–22]. Conversely, Drusano et al. [24,25] inoculated the thighs or lungs of mice with increasing doses of bacteria, and found a threshold in the initial bacterial density above which the infection grows uncontrolled (one could call this a critical bacterial concentration, CBC). A CBC is a natural outcome of a model in which the number of bacteria that can be killed by the maximum number of neutrophils approaches a maximum [12,13]. For instance, a model with a saturated killing rate, dBdt=rB−kNB/(1+B/h), in which a neutrophil can maximally kill kh bacteria per unit of time, has a CBC of B^=h(kN/r−1). Since this saturation model approaches the mass-action model at low bacterial densities, i.e., when B≪h, this model also has a CNC at low bacterial densities. Malka et al. [23] vary both neutrophil and bacterial densities in ex vivo experiments, and find evidence for a double saturation model, dBdt=rB−kNB/(1+B/hk+N/hN), which also obeys mass-action kinetics at low densities, but decreases the killing rate per neutrophil as a function of the neutrophil density (for which there is additional evidence [26]). However, for reasons of simplicity, we here adhere to the conventional model in which the killing rate only saturates with the bacterial density. 2.1 Results: Mouse models Leung et al. [26] developed a model allowing for a critical bacterial concentration, CBC, by also using the conventional saturated killing term, and in a second paper Roach et al. [10] showed that this model can explain how phage therapy can cure mice suffering from a rampant bacterial infection in the lungs by bringing the infection below the CBC. Because this model accounts for most of the in vivo observations during successful phage therapy in mice [10], we take their model as a starting point. The model defines phage-sensitive bacteria, S, phase-resistant bacteria, R, and bacteriophages, P, where N is the density of neutrophils, and B = S + R is the total density of bacteria. Bacteria grow logistically, with a net replication rate, r, have a carrying capacity, K, and a mutation rate, μ. Resistant bacteria may suffer from a fitness cost, ϕ≤1. Mutations are not reversible, i.e., sensitive bacteria mutate into resistant bacteria, but mutations of resistant bacteria back to sensitive strains are not considered in the model (other mutations are included in the net replication rate). Time units are per hour and densities are defined in terms of numbers per gram of tissue [10]. At low phage densities the infection rate obeys mass action kinetics, where individual susceptible bacteria are infected at a rate βP. At high phage densities, i.e., when P≫hP, sensitive bacteria are infected at a maximal rate βhP, and when P=hP the infection rate is half-maximal. The neutrophil killing rate was also defined by a conventional saturated killing term [26], saturating as a function of the bacterial density. At low bacterial densities the killing rate obeys mass action kinetics, with individual bacteria being killed at a rate kN, i.e., proportional to the density of neutrophils. At high bacterial densities individual bacteria are killed at a rate khkN/B (which is proportional to the effector:target ratio, i.e., N/B). At very high bacterial densities, i.e., when B≫khkN, the killing rate by neutrophils will therefore have a negligible effect, and bacterial densities will approach their carrying capacity, K. In Eq. (3) phages are cleared at a per capita rate dP (which is independent of their absorption to bacteria), and b defines a burst size, i.e., the number of phages produced due to the infection of a susceptible bacterium. In Eq. (4) the recruitment of neutrophils into the lungs was modeled by logistic growth at a maximum rate, rN, which saturated as a function of the bacterial density, B. The carrying capacity, Nmax, of this logistic equation defined the maximum density of neutrophils per gram of tissue, and their initial density (in the absence of bacteria) was defined as an independent parameter, N(0) (note that in the absence of bacteria dNdt=0). We continue by mildly modifying the model and re-estimating some of the parameters. First, as we later require an explicit equation for the infected cells, we define where dI is the rate at which infected cells burst. This is a minor change as making the quasi steady state assumption (QSSA), dIdt=0, gives and substituting this into dPdt in Eq. (5) we again retrieve Eq. (3), with a burst size b. Lytic phages typically take about half an hour to kill the bacteria they infected and produce a new generation of viral particles [27,28]. Having about two generations per hour, and a burst size of approximately b = 100 particles [10], an upper bound for the replication rate of phages would be that all b = 100 phages produced by an infected cell are infectious, and together rapidly infect a total of a hundred bacteria, which after half an hour each produce another generation of b = 100 phages. If this were a synchronous process during the first few generations, we can index the number of infected cells for each half hour, i.e., t=0,0.5,1,1.5,…, by the map and so on. Thus, starting with one infected cell, I0 = 1, and ignoring viral clearance, there would be I1=b2=104 phages after one hour. To translate this level of expansion into the differential equations we are using, we can solve for the initial replication rate, ρ, from I(t)=I(0)eρt, i.e., from I(1)=104=eρ giving an upper bound ρ=9.21 h–1. Note that this estimate of ρ is quite robust because we are taking a logarithm by solving ρ=lnb1/L, where b = 100 and L = 0.5 is the latent period in hours. For instance, if the latent period were to be 1 hour, we would estimate ρ=4.6 h–1. One could also allow for infected cell clearance, by considering I(1)=103, and obtain ρ=6.91 h–1. For a bacterial population at carrying capacity, S = K, the initial replication rate in the Roach et al. [10] model follows from Eq. (3), which at low phage densities would approach the unrealistically fast ρ=bβK−dP=5.4×104 h–1 (parameter values are provided in the caption of Fig 1). However, because the mice were treated by administering a large number of phages before the bacteria approached their carrying capacity (see Fig 1), the initial per capita replication rate of the phages remained reasonably low, i.e., in Eq. (3) P/hP≫1 and S≪K. Since we consider treatments of rampant human bacterial infections in the second half of this paper, where the bacteria have approached their carrying capacity, we limit the initial replication rate by modifying the model in two ways. First, we slow the initial replication of the phages by not making the QSSA performed in Eq. (6), which limits the initial replication rate by the time scale at which bacteria are lysed (dI=2 h–1 [27,28]). Second, because at high bacterial densities the infection rate should no longer be limited by bacteria, we modify the infection term into a double saturation function which naturally approaches hBβP when S→∞, hPβS when P→∞, and a mass-action term βSP when both S and P are small [29,30]. The model therefore changes into where Eq. (11) has remained the same as Eq. (2). The initial replication rate of this model can be computed by considering the initial phase of a small population of phages (P≪hP) during a rampant bacterial infection (S=K≫hB). In the Methods (Section 4.1) we derive that in mice the replication rate ρ=9.21 h–1 is approached when hB=9.63×106 CFU/g (which is similar to the other saturation constant, hP=1.5×107 PFU/g, in Eq. (9)). Thus, if we were to use this saturation constant value for hB in the double saturation function, Eq. (9), we are guaranteed to have a feasible initial phage growth rate in mice. The characteristic behavior of the Roach et al. [10] model. Panels A–C illustrate their previous results with a large bacterial infection, S(0)=7.4×107 colony forming units (CFU/g) and R(0) = 0. Panel A depicts an infection without phages. Panel B depicts the same infection treated with a large dose of phages, P(2)=7.4×108 plaque forming units per gram (PFU/g), given after two hours. Panel C depicts the failure of the same treatment when neutrophils cannot expand beyond their initial level. Red lines depict sensitive bacteria, S, blue lines resistant bacteria, R, purple lines total bacteria, B = S + R, green lines phages, P, and orange lines neutrophils, N. The horizontal dashed black lines depict the CBC threshold B^Nmax=6.7×107 (this CBC is absent from Panel C because neutrophils cannot expand). Panel D illustrates that in the absence of phages bacterial infections starting below the CBC grow initially, and can only be controlled after the neutrophils have expanded significantly, and that bacterial infections starting above the CBC grow uncontrolled to their logistic limit. Infections are started with S(0)=106,107 to S(0) = 108 CFU/g, and the line width of the curves identifies the three initial conditions. Roach et al. [10] parameters: r = 0.75 h–1, K = 1010 CFU/g, ϕ=0.9, μ=2.85×10−8, β=5.4×10−8 g/(PFU h), hP=1.5×107 PFU/g, b = 100, dP=0.07 h–1, k=8.2×10−8 g/(cell h), hk=4.1×107 CFU/g, rN=0.97 h–1, hN=107 CFU/g, N(0)=2.7×106 cells g–1 and Nmax=2.4×107 cells g–1. 2.1.1 Neutrophil recruitment. In the original model [26], the recruitment of neutrophils into the lungs was modeled by logistic growth at a maximum rate, rN, which saturated as a function of the bacterial density, B; see Eq. (4). The neutrophil equation was fitted to neutrophil dynamics data obtained from the lungs of mice stimulated with LPS [14], which suggested an initial density of N(0)=2.7×106 neutrophils/g, and maximal density of Nmax=2.4×107 cells/g (i.e., an almost 9-fold increase). We modify the neutrophil model for several reasons. First, recruitment of neutrophils into tissues actually consists of two phases, i.e., a very fast phase where circulating neutrophils leave the blood locally to enter an inflamed tissue, and (2) a slower phase where the number of circulating neutrophils increases due to an increased release of immature and mature neutrophils from the bone marrow [15,16]. The 24 h time scale of the Reutershan et al. [14] experiments, and the fact that the number of neutrophils in the lungs exceeded the initial number present in the circulation, implies that the data largely reflect an increased release from the bone marrow. Second, since in Eq. (4) dNdt=0 when B = 0, neutrophil densities fail to return to their initial density after a bacterial infection has been cleared. We therefore propose where σ is the normal production rate of circulating neutrophils by the bone marrow, and α≃8 would allow this to be increased approximately 9-fold at high bacterial densities. The neutrophil death rate dN allows the neutrophil density to return to the initial state N(0)=σ/dN, after clearing the bacterial infection. The maximum neutrophil density that is approached for large bacterial densities is Nmax=σ(1+α)/dN. Thus, we replace their logistic growth with Eq. (14) and adopt the previous parameters by defining, σ/dN=N(0) and α=Nmax/N(0)−1, such that we have the same initial and maximal density of neutrophils. The loss rate of the neutrophils, dN, defines the time scale at which the new Nmax is approached, and we have tuned dN and hN to have a similar increase of neutrophils during the first phase of the infection. Below we will adapt this model for fitting human data. The more mechanistic Eq. (14) has the additional advantage that we have reasonable estimates for dN in humans. 2.1.2 Neutrophil killing rates. The killing rates of the original model were based upon in vivo experiments in the lungs of mice in which bacterial expansion was studied, as a function of the number of bacteria inoculated [24]. The data revealed that initial bacterial densities smaller than B(0) = 107 colony forming units per gram (CFU/g) can be controlled by the neutrophil response, and that larger densities grow uncontrolled [24]. By describing the data with a model with a saturated killing term similar to the one in Eq. (1), i.e., where we combine the killing rate and the neutrophil density into κ=kN, we have a model allowing for such a threshold [13]. This threshold can be found from Eq. (15) by computing the bacterial density at which dBdt>0. Solving the full quadratic equation defined by setting dBdt=0,Leung et al. [26] and Marchi et al. [31] derive the CBC threshold B^=hk(κ/r−1) (the same threshold can also be found by considering bacterial densities way below the carrying capacity, i.e., by solving dBdt=B[r−κ/(1+B/hk)]=0 for B). Since κ=kN, this CBC depends on the neutrophil density, and because the model approaches two typical neutrophil densities, N(0) and Nmax, we also define two thresholds. The lower CBC threshold, B^N0=hk(kN0/r−1), defines bacterial infections that are immediately controlled, and need no inflammatory response, and the upper CBC threshold, B^Nmax=hk(kNmax/r−1), defines bacterial densities above which the maximum neutrophil response fails to control the infection. The two thresholds will be depicted as dashed or dotted lines in all figures. Drusano et al. [24] fitted data they collected on the rate of Pseudomonas aeruginosa killing by neutrophils in a murine model of pneumonia with a mathematical model, and some of the parameters they estimated were adopted by Roach et al. [10], i.e., r = 0.75 h–1 (a doubling time of 0.9 h) and κ=2 h–1. Assuming that this estimated total killing rate, κ, was due to a neutrophil density at its maximum, Leung et al. [26] subsequently estimated the mass action killing rate k=κ/Nmax [10]. By setting the saturation constant to hk=4.1×107 CFU/g, their model was given the property that bacterial infections starting at a density below B(0)≃107 CFU/g are controlled, whereas larger densities approach the carrying capacity (see Fig 1D). For these parameters individual neutrophils can maximally kill khk=3.4 CFU/g bacteria per hour, and the upper threshold is located at B^Nmax≃6.7×107 CFU/g. However, the combination of their bacterial replication rate, r, their neutrophil killing rate, k, and their normal neutrophil level, N(0)=2.6×106, does not allow small bacterial infections to be controlled without an inflammatory response, i.e., kN(0)/r < 1, implying that their lower threshold, B^N0, is negative, and that N(0) is below the CNC. To improve upon this, and to resolve issues with the identifiability of the parameters, we refit the Drusano et al. [24] data with somewhat simpler models in the Methods (Section 4.2). Since we obtain new parameter estimates from the same data, with a faster mass-action killing rate, k, and a lower saturation constant, hk, we obtain that small bacterial infections can be controlled without increasing the neutrophil density, and maintain the feature that large bacterial infections cannot be controlled. For these new parameter estimates the upper CBC becomes B^Nmax≃3.1×107 CFU/g (which seems reasonable because this is just 2-fold lower than the previous estimate [10]; see Table 2 in the Methods). 2.2 Results: Mouse experiments The most important contribution of the Roach et al. [10] paper was to demonstrate that the neutrophil immune response is essential to make phage therapy for Pseudomonas aeruginosa work in mice. Mice were infected intranasally with large doses of P. aeruginosa, and two hours later a large dose of a Pseudomonas phage was given intranasally. Time courses of bacteria and phages were followed over 72 hours. Importantly, the phage therapy was beneficial only when the mice had normal neutrophil numbers, and when these were able to respond to the infection [10]. We illustrate their results in Fig 1A–1C. A large bacterial infection that starts just above the CBC approaches carrying capacity in 10 h (Fig 1A). The same infection is controlled by a high dose of phages administered after two hours, by bringing it below the CBC within ten hours (Fig 1B). The same phage therapy fails when neutrophils cannot expand (Fig 1C), resulting in a chronic infection with phage-resistant bacteria. These model predictions were elegantly confirmed by experiments, where the failure of the therapy in Fig 1C was tested with MyD88-deficient hosts that lack immune cell activation and recruitment to the site of infection [10]. The fact that for the original parameters the model required expansion of the neutrophils to control even very small bacterial infections is illustrated in Fig 1D, where we depict a series of bacterial challenges, from S(0) = 104 to 108 CFU/g, that all grow initially, and all evoke an increase in the neutrophil densities. Fig 1D also confirms that challenges above S(0) = 107 grow uncontrolled (despite the fact that this is below the CBC located at B^Nmax=6.7×107 CFU/g; see the horizontal black line). We have modified the Roach et al. [10] model by replacing the logistic growth term for the increase of the neutrophils by a more natural model having a source from the bone marrow (see Eq. (14)), by making the infection rate a saturation function of both the phages and the bacteria, by not making the QSSA for the infected bacteria, by increasing the killing rate, k, and by decreasing the saturation constant, hk, such that bacterial infections smaller than B(0) = 106 CFU/g can be cleared without requiring increased production of neutrophils in the bone marrow. Because we have tuned the parameters of the new neutrophil model to that of the previous model, we observe very similar establishment of a chronic infection with a large challenge of bacteria (Fig 2A). The new model has two CBCs, B^N0=106 CFU/g and B^Nmax=3.1×107 CFU/g, for the normal and the maximum number of neutrophils, respectively (see Table 2). The effect of the two thresholds is illustrated in Fig 2B, where we depict a similar series of bacterial challenges, from B(0) = 105 to 108 CFU/g, in the absence of phages. Infections starting below B^N0=106 CFU/g decline immediately, and infections starting with B(0) = 107 CFU/g or more grow uncontrolled (similar to the original model). Similar behavior of the new model given by Eqs. (10)–(13) and (14). Panel A is like Fig 1A, an infection without phages. Panel B is like Fig 1D with infections starting with S(0)=105,106,107 to S(0) = 108 CFU/g. Infections starting below B(0)=B^N0 decline monotonically, and those above S(0)=9×106 CFU/g grow uncontrolled (in the absence of phages). Panel C is like Fig 1B, a controlled infection due to a large dose of phages, P(2)=7.4×108 PFU/g, given after two hours. Panel D depicts a controlled infection due to the same large dose of phages, in the absence of an increase in neutrophils (α=0). Panel E depicts the evolution of resistance (like Fig 1C) obtained by decreasing the infection rate 2-fold, and starting with R(0)=5000 CFU/g resistant bacteria. Panel F is like Panel C, a controlled infection, but due to a small dose of phages, P(2) = 1 PFU/g, given after two hours. The horizontal black dashed lines represent the low and the high CBC, B^N0 and B^Nmax, respectively (the high CBC is absent when α=0). Colors as in Fig 1, S: red, R: blue, B: purple, P: green, and N: orange. New parameters: σ=2.25×105 cells h–1, dI=2 h–1, dN=0.0833 h–1 (i.e., 0.5 d–1), r = 0.3 h–1, k=1.5×10−7g/(cell h), α=7.9, dN=1/12 h–1, hB=9.63×106 CFU/g, hN=105 CFU/g, and hk=2.86×106 CFU/g. We obtain a very similar successful phage therapy with a large dose of phages after two hours (Fig 2C). A novel feature of the new model is that phage therapy can also work when the bacterial densities drop below the lower CBC; see Fig 2D, where we have set α=0 to prevent expansion of the neutrophils. Increased production of neutrophils is therefore not strictly required to make phage therapy work. The new model can also account for a similar failure of phage therapy, and the evolution of resistance (Fig 2E), when we increase the initial number of resistant bacteria and decrease the infection rate. Finally, we observe in the new model that the initial dose at which the phages are administered hardly matters because they expand vigorously in the presence of a large density of susceptible bacteria (compare Fig 2C with Fig 2F), despite the fact that we have bounded the maximal replication rate by not making the QSSA dIdt=0, and by using the new infection term of Eq. (9). Likewise, replacing the single dose of phages with a small influx starting after two hours, representing the intravenous administration used to treat patients [2,4,6], yields a time course that is very similar to that shown in Fig 2F (see below). What have we learned from modeling phage therapy in mice? First, we confirm that the dose at which phages need to be administered to successfully control a bacterial infection hardly matters [26]. If the phages are able to infect the bacteria they will expand rapidly in a few generations, even in a model where we limit the replication rate of phages to ρ=9.21 h–1. Clinically this is interesting because it is important that the phage preparations are sterile, and have endotoxin levels below FDA-approved levels [1,2], which is easier to achieve with small than with large doses. Second, by re-estimating the rate at which neutrophils kill bacteria, we now have two thresholds in the model, and successful phage therapy no longer requires an increased neutrophil production since the bacterial infection will also become controlled when bacterial densities drop below the level that can be controlled by normal (non-inflamed) neutrophil densities. This could also be important clinically because it is not known how long neutrophil responses remain elevated during chronic bacterial infections. As a next step we develop this mouse model further by applying it to the human situation where patients are treated with cocktails of phages that the bacteria can independently evolve resistance to. 2.3 Results: A general human model Because patients are treated with cocktails of phages, and bacteria independently evolve resistance to all of them, we have to extend the model by allowing for several strains of bacteria, and define the mutation rates at which these strains evolve from one another. Because patients are infused with phages we also add a time dependent source, V(t), to the ODE for the phages. We also add a strain-specific loss of bacteria to model treatment with antibiotics, because patients are often treated with various antibiotics, as sensitive and resistant strains frequently differ in their sensitivity to antibiotics [2]. For instance, Chan et al. [5] used phages binding the membrane proteins of a multi-drug efflux system, such that bacteria developing phage-resistance by down regulating these receptors became more sensitive to antibiotics. Because the concentration of phages in human serum can decline quite rapidly [4], we estimate the loss rate of phages, dP, by modeling the pharmacokinetics of the phages in serum after intravenous administration, as measured by Schooley et al. [4] in the Methods (Section 4.5). Fitting simple models to that data we estimate that in humans dP=3.5 h–1 (see Fig 9). An alternative would be to allow for a loss of phages by absorption to bacteria, and/or a loss of phages due to the (slow) development of phage-specific antibody responses by the host [2]. Finally, since one of the potential mechanisms underlying the success of phage therapy is a loss of bacterial fitness due to their evolution of phage resistance, we extend the logistic bacterial growth model by including a density independent death rate, dB, of the bacteria, i.e., dBdt=ϕrB(1−B/K)−dBB. This allows the steady state of a bacterial strain, B¯=K(1−dB/(ϕr)), to depend on its fitness ϕ. In the previous model the steady state of bacteria in the absence of phages and neutrophils was defined as B¯=K, and it did not depend on their fitness ϕ. The Roach et al. [10] model was recently extended to study treatment with two phages [31], and since we aim to model bacterial strains resistant to several ‘functionally different’ groups of phages [32], we generalize upon this by using a conventional bit string notation, where the strings ‘000’, ‘001’, ‘010’, ‘011’, ‘100’ have 0, 1, 2, 3, and 4 as their decimal interpretation, respectively. Also ordering the phage groups from right to left in the string, we define bacteria B1 with bit string s1 = 00 to be susceptible to phage groups P1 and P2, bacteria B2 with the bit string s2 = 01 to be resistant to group one only, bacteria B3 with the bit string s3 = 10 to be resistant to group two only, and bacteria B4 with bit string s3 = 11 to be resistant to both groups P1 and P2 (i.e., the bacterial strain number i is one plus the decimal interpretation of its string). A model with nP different phage groups, each requiring a unique resistance mechanism, therefore needs bit strings of nP bits defining nB=2nP different types of bacteria, and nPnB types of infected cells, and a nP×nB susceptibility matrix Mij defining all possible routes of infection. Assuming that each phage group requires it own unique resistance mechanism, evolving resistance to them requires different ‘mutations’, i.e., deletions, insertions and point mutations, with different mutation rates. We therefore generalize the single mutation rate, μ, into a vector, μi, where i=1,2,…,nP. Since each bacterial strain i can evolve from any other strain j we combine these mutations into a matrix, μ^ij. To keep the model simple we also only consider ‘gain of resistance’ mutations [10,26], i.e., the upper diagonal of the mutation matrix is zero. For example, a fully susceptible B00 strain will mutate to B01 and B10 strains, at rates μ1 and μ2, respectively, and into strain B11 at a rate μ1μ2. For these four bacterial strains the mutation matrix would look like where each column defines the rate at which Bi mutates into the Bjs, which is a loss term for Bi and a gain for a Bj (see below). Following the original model developed for mice [10,26], mutations occur when bacteria divide. In chronically infected patients resistant bacteria are expected to be present before phages are introduced, as bacteria have evolved rapid resistance mechanisms over millions of years of co-evolution with phages [33]. The fitness, ϕi, of bacterial strain i can be estimated from data (see Methods; Section 4.4), or can be defined as a weighted sum (or a product) of the costs associated with each resistance mechanism, e.g., ϕi=∑j=1nPcjsij, where cj is the cost associated with mutation, μj, and si is the bit string of bacterial strain i, and j is the phage number (and the position in the string when counting from right to left). Finally, bacteria may become more sensitive to killing by neutrophils when they become resistant by modifying their outer membrane, [7]. We allow for this by giving each bacterial strain its own rate, ki, of being killed by neutrophils. Since we now consider several phage groups we need to reconsider the saturation in the infection term. As different phages, Pi, may use the same receptor molecule on the bacteria to enter and infect, we mechanistically derive a high-dimensional infection term by ordering the phages into groups using the same receptor, and deriving a saturated infection term for all phages within one group (see the cartoon (Fig 7) and the full derivation in the Methods; Section 4.3). Due to its complexity we propose to lump all phages within one functional group into one “super phage” with an “average” infection rate, βi, and saturation constant, hP. This has the additional advantage that the model can be written with one phage-resistance mechanism per group, i.e., if the bacteria down regulate the receptor used by group i, all phages in that group would be affected. Thanks to this simplification, the form of the infection term remains similar to the one used previously [10]. Summarizing, we consider a population of sensitive and resistant bacteria, B=∑iBi, growing logistically. Bacteria, Bi, are infected by different groups of phages, Pj, to form infected cells, Iij, which burst and produce more phages, Pj. Bacteria are killed by a population of neutrophils, N, and the rate of killing is half-maximal when B=hk. Bacteria can only be infected by the phage groups they are susceptible to, which is summarized in the susceptibility matrix Mij defining all possible routes of infection. All of this can be formalized into the following model: for i=1,2,…,nB and j=1,2,…,nP, and where nP defines the number of phage groups. The vector βj defines the infectivity of phage strain j, Mij defines which bacterial strain is resistant (Mij=0) or sensitive (Mij=1) to what phage, and nA is the number of potentially effective antibiotics a patient is treated with, where Aj is a constant defining the concentration of the jth antibiotic, and Sij defines the sensitivity of bacterial strain i for antibiotic j. Eq. (19) for the phages is extended with a time-dependent and phage-specific intravenous infusion rate, Vj(t), and a mass-action absorption to all bacteria at a rate a (as phages can bind bacteria that they cannot infect). The latter is a simplification when bacteria down-regulate the receptor molecule that specific phages use for binding and entering as a resistance mechanism. 2.3.1 Human parameters. The value and the biological interpretation of all parameters is specified in the Methods (Table 4). The original model was parametrized on mouse data per gram of tissue, and since bacterial and phage densities per gram of tissue are probably similar in mice and men, we use most parameters of the mouse model for the human situation (except when human estimates are available). Since patients with rampant bacterial infections typically have several tissues infected with bacteria, the neutrophil equation of the model can best be interpreted as the number of neutrophils in the circulation and peripheral tissues. In the absence of inflammation we define a normal value, N(0), and we allow this to be increased 10-fold at high bacterial densities, by setting α=9. Because we use the faster loss rate of phages that we estimated for humans, dP=3.5 h–1 (see Methods; Section 4.5), the phage densities now decline to much lower levels when the bacteria become resistant or disappear due to successful therapy. To allow a population of phages to go extinct (and not grow back from unrealistic ‘atto-fox’ levels when sensitive bacteria recover [34]), we multiply the infection rate with a steep Hill function, i.e., we use a ‘truncated’ infection rate, βj/(1+[θ/Pj]5) (note that Marchi et al. [35] use a similar function to limit the diffusion of phages in a spatial model). Thus, the infection rate is halved when Pj=θ, it rapidly approaches zero when Pj<θ, and it rapidly approaches βj when Pj>θ. Since we are modeling one gram of tissue, and one PFU should approximately correspond to a single phage, we set θ=10−4 PFU/g to allow phages to go extinct when there is less than one particle in a total body weight of about 10 kg of relevant tissue (setting θ either 10-fold higher or lower does not affect our main results). We estimate the infection rate of the various phages by the successfulness of the treatment. Phage therapy will be successful when it reduces the bacterial density several orders of magnitude, from its carrying capacity to breach a CBC below which neutrophils can eliminate the infection (this CBC may depend on the bacterial strain when strains experience different killing rates, ki). The bacteria will only decrease when the force of infection, F, exceeds the maximal net replication rate, r−dB. At high phage densities, i.e., Pi≫hP, the maximum infection rate per phage group approaches βihP (summing up to a force of infection ∑inPβihP for all phages). Since the phages rapidly approach high densities, we define a critical infection rate above which one phage group can control the bacteria, by solving βcrithP=r−dB, i.e., βcrit=(r−dB)/hP=2×10−8 g per PFU per hour. To maximize the initial replication rate of phages to ρ=9.21 h–1, we now substitute the human parameters from Table 4 into Eq. (24) and estimate that hB=3.56×107 CPU/g (for β=βcrit). 2.3.2 Single phage group. To explore the behavior of this new ‘human’ model we first simulate a simplified phage therapy with a single phage group, while varying the infection rate, β, and the mutation rate, μ. We consider a bacterial infection at steady state, with a corresponding increased number of neutrophils, and start with an arbitrary small number of phages, P1(0) = 1 PFU/g, on day zero (see Fig 3). Because bacterial densities are high the phages approach their maximum replication rate of ρ=9.21 h–1, and although their replication rate slows down when the phage densities approach hP, they very rapidly expand to densities exceeding 1010 PFU/g (Fig 3). Because phages replicate so rapidly when bacteria are abundant, this confirms that the initial dose hardly matters. Exploring the human model with a single phage. Sensitive bacteria are now denoted as B0 and resistant bacteria as B1 (instead of S and R in the previous model). In Panels A & B, C & D, and E & F, we set β=10−8, β=2×10−8, and β=3×10−8 g per PFU per hour, respectively, i.e., below, at, and above βcrit. In Panels A and C the mutation rate is set to zero, in Panels B, D and E we have kept the μ=2.85×10−8 h–1 that was used by Roach et al. [10], and in Panel F the mutation rate is increased 10-fold. In the absence of resistance, the phage clears the bacterial infection when β≥βcrit; compare Panels A and C. In the presence of resistance, an infection rate larger than βcrit need not be sufficient, because the resistant B1 bacteria may breach the upper CBC threshold (B^Nmax); see Panel D and Panel F. In Panel B the phages do not go extinct: the system slowly approaches a steady state dominated by a large population of resistant bacteria, with neutrophils at their maximum density, and the phages and the sensitive bacteria approach this steady state in a dampened oscillation (see a longer simulation in S2A Fig). In Panels D and F the phages are declining while the sensitive bacteria, B0, are slowly recovering. Here the phages do go extinct, and the sensitive bacteria slowly take over from the resistant bacteria due to their 10% higher fitness (see S2B and S2C Fig). For better visibility the neutrophil concentration is divided by N(0) (i.e., N is scaled between one and ten). Novel parameter values: ϕ1=1, ϕ2=0.9, r = 0.31 h–1, dB=0.01 h–1, dP=3.5 h–1, hB=3.56×107 CFU/g, k1=k2=0.405 h–1, a = 0, and μ^=(0 0μ 0); see Table 4 in the Methods. Phage therapies with phages having a low infection rate (i.e., β<βcrit) fail because the sensitive bacteria just approach a somewhat lower level (see Fig 3A; where there are no resistant bacteria because we set μ=0), or only have a similar small transient effect because the resistant bacteria rapidly take over, and the bacterial infection returns to a similar high steady state (Fig 3B). Since the phages are declining after most of the bacteria have become resistant, the sensitive bacteria start to recover slowly (from day 7 onward in Fig 3B). Ultimately, this also allows the phages to recover, and the system slowly approaches a steady state dominated by resistant bacteria and high neutrophil levels, and about 107 CFU/g sensitive bacteria maintaining a population of about 105 PFU/g phages (S2A Fig). In the absence of resistance, phages with a higher infection rate of βcrit suffice to control the bacterial infection (Fig 3C), but in the presence of resistance this infection rate need not be sufficient (Fig 3D), because the bacteria rapidly become resistant and the phages go extinct (S2B Fig). Disturbingly, the ‘resistant’ steady state is approached faster when the phages are more infectious, because phages having a higher infection rate lead to a steeper decline of the bacteria, which reduces the logistic competition, and allows the resistant bacteria to replicate faster (compare the total bacterial load, B, in Fig 3B with that in Fig 3D). Increasing the infection rate further allows the phages to clear the bacterial infection in the presence of resistance (Fig 3E), because more infectious phages more rapidly deplete the sensitive bacteria, which here brings the total bacterial density below the upper threshold CBC, B^Nmax, before the resistant bacteria have taken over. Note that resistant bacteria start declining when the total, B, breaches the upper threshold, that the phages disappear afterwards quite rapidly due to their relatively high turnover, and that the neutrophils return to their normal steady state (Fig 3C and 3E). However, a therapy with the same more infectious phages fails when the bacteria more easily become resistant (Fig 3F), because it then takes less time for the resistant bacteria to breach the upper threshold CBC. Summarizing, whether or not phages can control a rampant bacterial infection depends both on their infection rate and on the rate by which the bacteria develop resistance, as phage therapy will fail as soon as the resistant bacteria breach the upper CBC. Although it should always be better to treat with highly infectious phages, an important new lesson learned is that resistant bacteria expand more quickly when the phages are more infectious. Having a basic understanding of the new model we next use it to simulate phage therapy with two phages. 2.3.3 Two phage groups. Consider a phage therapy consisting of two phages (or two cocktails of different phage groups) that are given sequentially. For simplicity, we consider phage groups with identical infection rates, and bacterial strains with identical mutation rates, killing rates, and fitness costs (see Fig 4). Since a therapy with a single phage failed because of rapid development of resistance (Fig 3C), we add on a second phage group on day five (Fig 4A), or day two of the treatment (Fig 4B and 4C). The former fails because the frequency of bacteria resistant to both phage groups (B11 denoted by orange lines), markedly increased when the bacterial strains resistant to the first group (B01 denoted by blue lines) took over. This is a natural result because fully resistant bacteria are just one mutational step away from bacteria resistant to the first cocktail, and two steps away from susceptible bacteria (see Eq. (16)). One can prevent double-resistant bacteria to breach the upper CBC by giving the second treatment earlier (e.g., at day two; Fig 4B), which makes the combined treatment successful. Obviously, this timing depends on the initial frequencies of resistant bacteria (see Fig 4C where we increased the mutation rates 10-fold), and the optimal treatment would be to give both phage groups at the same time. Simulated phage therapy with two phages administered sequentially. The second phage is either given fairly late (at day five (Panels A and D)) or early (at day two: Panels B, C, E & F). Sensitive bacteria are now denoted as B00, and the 3 types of resistant bacteria as B01,B10 or B11, respectively. Both phages have the critical infection rate β=2×10−8 g/(PFU h). In the upper row the phages are given as a small single shot at day 0, 2 or 5, i.e., P1(0) = 1 and P2(2) = 1 or P2(5) = 1 PFU/g (see the color-matched arrows at the bottom of the graph) while in the lower row the treatment is given in the form of a continuous infusion, i.e., V1(t)=V¯ at all time points, and V2(t)=0 at the start until we set V2(t)=V¯ when t > 5 days (Panel D) or t > 2 days (Panels E & F); see the color-matched horizontal lines at bottom of the graph. The first five days in Panel (A), and the first two days in Panel (B) are therefore identical to the unsuccessful therapy shown in Fig 3D. Adding a second therapy at day five fails to prevent the evolution of resistance (Panels A & D), starting earlier makes the same therapy successful (Panels B & D), but a 10-fold higher mutation rate again leads to failure due to the evolution of resistance (Panels C & F). Importantly, there is hardly any difference in the total bacterial load, B, between single shots of the phages (Panels A–C), and a therapy comprised of continuous infusion (Panels D–F). Phages do not go extinct in Panel (C), and not when they are administered by continuous infusion (see S3C–S3F Fig). Novel parameter values: ϕ1=1, ϕ2=ϕ3=0.9, ϕ4=0.8, V¯=103 and μ1=μ2=2.85×10−8 h–1 in the matrix shown in Eq. (16). This resembles the ‘hit hard and early’ concept from antiretroviral therapy for HIV [18], where combinations of drugs requiring different escape mutations increases the genetic barrier to full resistance. Similarly, experiments with phage cocktails containing various combinations of functionally different phages, demonstrate that increasing the functional diversity markedly increases the efficacy of the phage therapy [32]. A major difference between antiretroviral therapy and phage therapy is that phages are replicating, and that phage therapy therefore relies much less on the dosing schedule. This is illustrated in the bottom panels of Fig 4 where we simulate a treatment consisting of a continuous infusion of the phages (rather than a small single ‘shot’ at a particular day). The high similarity in the total bacterial load, B, between the panels in the upper and lower row of Fig 4 confirms that dosing makes little difference. Continuous infusion only prevents the phages from going extinct when the availability of sensitive bacteria declines (see S3 Fig). The most important new lesson learned is that phage therapies should be given as early as possible to slow down the evolution of full resistance. Continuous infusion may reduce the reversal of resistance in a bacterial population, but even this would not require high doses as phages start replicating as soon as susceptible bacteria reappear. 2.4 Results: The Patterson case The first spectacular case of a patient recovering from a rampant bacterial infection with multidrug resistant (MDR) Acinetobacter baumannii after the administration of various cocktails of phages was described as the ‘Patterson case’ by Schooley et al. [4] and by Strathdee et al. [17]. Phage therapy was initiated about a 100 days after the initial infection while the patient was in coma, and although the bacteria developed resistance to all phages in the first two cocktails on a time scale of days, the patient woke up and started recovering [4,7,17]. Our aim here is to model this recovery during the first ten days of treatment. The first two days of treatment used a cocktail of phages called ΦPC, that was given via the catheters draining the intra-abdominal abscess cavities. It remained unclear whether or not the phages were able to replicate in this fairly acidic environment, and the patient did not improve. A second cocktail consisting of four other phages (ΦIV) became available on day three, and because the two days of ΦPC treatment seemed to have little effect, the treatment was switched to dosing intravenously. After these four days of phage therapy the patient awoke. The therapy was stopped the next day because of clinical problems, and as soon as this was diagnosed to be due to an unrelated bacterial infection, the treatment was reinstalled two days later, and continued for more than a 100 days [4]. After a week, minocycline was added to the treatment, but because the bacteria later turned out to be resistant to minocycline [7] we do not include any antibiotics in our model (i.e., we set c = 0 in Eq. (17)). A third cocktail was started on day 113, and because we only model the first 10 days during which the patient recovered, this is also not considered. Both the intra-abdominal and the intravenous infusion is accounted for with the V parameter in Eq. (19), and this particular time schedule of the treatments is implemented by setting Vi(t)=0 on day 1 and 2 for the phages in the ΦPC cocktail, and on days 5 and 6 for both cocktails. The genomes of the phages were sequenced and this revealed that they were all closely related to T4-like myophages, and that three out of the four phages in the ΦPC cocktail were actually identical [7]. Thus, the first cocktail consisted of two related phages called Maestro and AC4, while the second ΦIV cocktail consisted of four related phages called Navy1, Navy4, Navy71 and Navy97 (see Table 1). T4-like myophages use long tail fibers to bind to the phage’s receptor on the cell surface, and tail fiber analysis showed the myophages used in the cocktails had two different types of tail fibers, with Maestro, AC4, and Navy71 belonging to one group, and Navy1, Navy4, and Navy97 belonging to the other cluster. This agreed well with the observation that at day two, the bacteria became resistant to all three phages in the first cluster, and remained sensitive to the three in the second cluster (Table 2 in Liu et al. [7]). Importantly, this suggests that although the phages in the first ΦPC cocktail had little clinical effect, they were able to replicate and select for resistance. On day six, i.e., after four days of treatment, bacteria were resistant to all phages [7]. 10.1371/journal.pcbi.1014408.t001Table 1Composition of the two cocktails used by Schooley et al. [4]. The 6 phages can be classified in two resistance groups by their tail fibers [7], which is indicated in the rows.Cocktail:ΦPCΦIVResistance cluster 1Maestro & AC4Navy71Resistance cluster 2—Navy1, Navy4 & Navy 97 Composition of the two cocktails used by Schooley et al. [4]. The 6 phages can be classified in two resistance groups by their tail fibers [7], which is indicated in the rows. 10.1371/journal.pcbi.1014408.t002Table 2The killing parameters of Roach et al. [10], and new estimates obtained by refitting the Drusano et al. [24] experiments in mice using Eq. (15).r (h–1)κ (h–1)k (g/(h cell))hk (CFU/g) khk B^N0 B^Nmax Roach et al. [10]0.751.97 8.2×10−8 4.1×107 3.4— 6.7×107 New estimates0.30.5 1.5×10−7 2.86×106 0.43106 3.1×107 The killing parameters of Roach et al. [10], and new estimates obtained by refitting the Drusano et al. [24] experiments in mice using Eq. (15). The fact that resistance comes in two clusters suggests that this can be modeled with two phage groups. Since Navy71, belonging to the first resistance cluster, was only present in the second cocktail (see Table 1), which was administered after the bacteria were already resistant to it, we did not model it separately. Because resistance to Navy1, Navy4 and Navy97 evolved simultaneously [7], we also treat the second cocktail as one uniform group of phages. Since bacteria became resistant to the phages in the second cluster by a common 6 base pair deletion in genes predicted to be involved in the synthesis of their capsule [7], a ‘mutation’ which led to the loss of the bacterial capsule should also make them resistant to the phages in first cluster. Thus, we ignore the B10 bacteria and consider only two ‘mutation’ rates in Eq. (16). The sequencing data confirmed that the resistant strains evolved from a common ancestor [7], and in the model we let sensitive bacteria develop resistance to cluster one at a rate μ^12=μ1, and let B00 and B01 strains shed their capsule at a rate μ^14=μ^24=μ2. All other elements in the mutation matrix of Eq. (16) are zero (i.e., the B10 strain remains absent), and we will vary μ1 and μ2 to match the data. In the first paper on the Patterson case [4], the development of phage-resistance was measured by culturing bacterial isolates from the patient in the presence and absence of the phages in the cocktails. We fit these in vitro data in the Methods (Section 4.4) to estimate the relative growth rates of the bacterial strains resistant to the first and second cocktail. The estimated fitness of pretreatment bacteria resistant to the first cocktail was estimated to be ϕ2=0.74, and their initial abundance was estimated to be high (21%) (see Table 3). Culturing bacterial isolates taken at day six revealed that the estimated fitness of bacteria resistant to both cocktails is about ϕ4=0.33 (see Table 3), with an estimated abundance of about 30% (at day six). The mutation rate leading to the B01 strain resistant to the first cocktail was set to μ1=0.05 to match the high initial abundance in the respiration experiments. The mutation rate leading to resistance to both cocktails was set to a 20-fold lower value to account for the later appearance of the B11 strains (that are also resistant to the first cocktail). Note that these mutation rates are orders of magnitude higher than those used above. 10.1371/journal.pcbi.1014408.t003Table 3Fitting the respiration data of Schooley et al. [4]. The values between the brackets denote standard errors computed from the Hessian matrix by the FME package [41]. Common parameters were similar for the pretreatment and the day-six experiments, i.e., dB=0.042 and e = 0.0036, and dB=0.051 and e = 0.0029, respectively.PretreatmentDay sixParameterNo phagesCocktail ΦPCNo phagesCocktail ΦIVr (h–1)0.58 (±0.03)0.43 (±0.01)0.51 (±0.01)0.22 (±0.02)B(0)4.49 (±1.05)1.01 (±0.15)6.56 (±0.68)1.95 (±0.51)Growth rate ρ0.43/0.58 = 0.740.22/0.51 = 0.44Fitness ϕi ϕ2=0.74 ϕ4=0.74×0.44=0.33 Frequency1.01/4.49 = 0.211.95/6.56 = 0.30 Fitting the respiration data of Schooley et al. [4]. The values between the brackets denote standard errors computed from the Hessian matrix by the FME package [41]. Common parameters were similar for the pretreatment and the day-six experiments, i.e., dB=0.042 and e = 0.0036, and dB=0.051 and e = 0.0029, respectively. At the estimated abundance of about 20% resistant B01 strains in the pretreatment bacterial population, the resistant strains are way above the upper CBC (see day zero in Fig 5A), and hence they rapidly take over during the first two days of treatment (whatever the infection rate of the phages). Because the first cocktail was administered into the fairly acidic environment of the abdominal abscess cavities, we assigned it an infection rate below the critical infection rate, i.e., β1=βcrit/2, and nevertheless observe that this unsuccessful cocktail does select for resistance in the bacteria. After two days of treatment the bacterial population changed from 81% susceptible B00 strains, to 98% B01 and 1% B11 strains (see day two in Fig 5A). Because this 1% abundance of fully resistant B11 strains also exceed the upper CBC, treatment with the second cocktail would also be unsuccessful (whatever the infection rate of the phages); see Fig 5A. Upon the introduction of the second ΦIV cocktail, the dominant B01 strains immediately decline, which accelerates the replication rate of the fully resistant B11 strains, that rapidly become dominant. Thus, the first cocktail markedly increases the resistance to the second cocktail even before this ΦIV cocktail is administered. Note that increasing the infectivity of the first cocktail would only speed up the development of full resistance, without making it successful. The Schooley et al. [4] patient modeled with two phage groups. The first cocktail is given on day 1 (i.e., at t = 0) and a second on day 3 (i.e., at t = 2); see the color-matched horizontal lines at the bottom of the graph. In Panel A the infection rate of cocktails is set to β1=βcrit/2=10−8 and β2=2βcrit=4×10−8 g/(PFU h), which leads to failure of the treatment because at the estimated pretreatment level of resistance both types of resistant bacteria are above the upper CBC (with μ1=0.05 and μ2=μ1/20=0.0025). In Panel B we allow for a 10-fold faster killing rate of the fully resistant B11 bacteria, which clears the infection. In Panel C we decrease the second mutation rate 5-fold (to μ2=μ1/100=0.0005), which keeps the fully resistant bacteria below the upper CBC, which also makes the second cocktail successful. In Panel D the combination therapy of Panel A becomes successful when both cocktails are given at the same time, because the fully resistant bacteria have less time to expand. We use the same parameters as in Fig 3 (see Table 4), but use the estimated abundances and growth rates to set the mutation rates μ1 and μ2, and the fitness values ϕ1=1, ϕ2=0.74 and ϕ4=0.33 (see Table 3). For better visibility the neutrophil concentration is divided by N(0) (i.e., N is scaled between one and ten). Since Fig 5A is not in agreement with the data as the second ΦIV cocktail was successful, we consider two possibilities. First, the fully resistant B11 strains are much more sensitive to being killed by neutrophils because they have a deficient capsule [7]. In Fig 5B we therefore increase the killing rate of B11 bacteria, k4, by a factor 10, which is sufficient to explain the success of the treatment. Increasing the killing rate increases the upper CBC for fully resistant bacteria, which allows them to be controlled by neutrophils as soon as the total density of partially resistant bacteria has declined sufficiently by infection with the ΦIV phages (see Fig 5B). Note that the increased killing rate also decreases the initial abundance of fully resistant bacteria (see day zero in Fig 5B), which would be helpful if their abundance were to remain below the upper CBC. A second possibility is that the initial abundances estimated from the respiration data are overestimated because (1) the relationship between the true bacterial density and the observed respiration need not be linear [36], (2) the background respiration that we subtracted from the data was large compared to the initial respiration of the bacteria, and (3) we do not have an estimate for the pretreatment abundance of the fully resistant bacteria because they hardly replicated in the respiration experiments [4]. Explaining the success of the second cocktail just requires a much lower abundance of the fully resistant bacteria, i.e., by setting μ2=μ1/100, the fully resistant bacteria remain below the upper CBC, even after their rapid expansion during the first two days of treatment. This also suffices to explain the clinical success of the second cocktail (Fig 5C). Because the resistance to the second cocktail markedly increases during the treatment with the first cocktail, we improve the unsuccessful treatment of Fig 5A by giving both cocktails at the same time (see Fig 5D). Simultaneous administration makes the phage therapy successful because the fully resistant strain is not given the opportunity to expand during the treatment with the first cocktail. Summarizing, whether or not phage therapy can clear a rampant bacterial infection requires sufficiently low pretreatment levels of resistance, and administration of all phages at the same time may prevent the rapid expansion of fully resistant bacteria. Experiments comparing the efficacy of simultaneous and sequential treatments confirm that adding phases simultaneously tends to be the most successful [36]. Our modeling adds the disturbing insight that increasing the infection rate in an earlier cocktail leads to a more rapid development of resistance to the next one. 2.5 Results: Functional diversity of the phages In experiments, increasing the number of functionally different phages in cocktails added to bacterial cultures increased the efficacy with which the cocktails suppressed bacterial growth in a saturating manner [32]. In our model, the major advantage of an immediate parallel treatment with several functionally different phage groups is that the initial density of the fully resistant bacteria will still be at the pretreatment level, and that for each additional phage group requiring another resistance mutation, the expected pretreatment concentration will drop by its mutation rate. The lower the pretreatment concentration the longer it will take a fully resistant strain to breach the upper CBC, and hence the more likely that the treatment works by bringing the total bacterial load below the upper CBC. To address the question whether it is better to improve the infection rate of the phages, or to increase the diversity of a cocktail of phage groups (that are all given immediately), we performed a parameter sweep increasing β¯ and nP (see Fig 6). Assuming identical mutation rates μ¯ for all phage groups, we use Eq. (16) to define the full mutation matrix μ^. We vary the mutation rate in the range suggested by the cultured isolates, i.e., μ¯=0.0025,0.005 and 0.01 in Fig 6A–6C, respectively, and we vary the common infection rate β¯ 128-fold by increasing it in 2-fold steps. This reveals that increasing the infection rate above βcrit hardly improves the successfulness of the therapy. Note that we do not adjust hB when β¯ changes, and that phages with an infection rate exceeding βcrit, will therefore initially replicate faster than ρ=9.21 h–1. This does not affect our result that increasing the infection rate hardly matters, because these initial phage replication rates allow for even better control. Even extremely infectious phages require a minimum diversity of phage groups to make phage therapy successful at these pretreatment resistance levels. Thus selecting for functionally different phages, requiring different resistance mechanisms, is much more important than optimizing individual phages. This is in good agreement with the experimental results of Wright et al. [32]. Parameter sweep, varying β¯ and the number of phage groups. In Panel A, B and C we set μ¯=0.0025,0.005 and 0.01, such that about 5%, 10% and 20% of the pretreatment bacteria are resistant to one of the phage groups, respectively. Each heat map depicts a parameter sweep, varying β¯ 128-fold, and the number of phage groups from 1 to 6 (all start at t = 0). The fitness of the bacteria declines proportional to the number of resistance mutations of the bacterial strain, i.e., ϕi=1−0.05∑j=1nPsij, where si is the bit string of strain i and j denotes the position in the string. Green boxes denote successful therapy and red boxes depict failure. Whenever β¯>βcrit, increasing the infection rate hardly affects the successfulness of the therapy. Even at very high infection rates, β¯>100βcrit (where βcrit=2×10−8 g/(PFU h), the pretreatment resistance levels, as set by the mutation rates, require a minimum number of phage groups, i.e., 1, 2 and 3, for μ¯0=0.0025,0.005 and 0.01, respectively. 3 Discussion Having developed a novel mathematical model for phage therapy, and after confirming that this model can successfully describe clinical data, we now have an in silico tool enabling us to compare potential scenarios for treatment. Most importantly we have observed in our models that during the first days of successful treatment the replication rate of (partially) resistant bacterial strains markedly increases because they are released from the severe competition with the previously dominant pool of sensitive bacteria. Whether or not a therapy will be successful depends on time it takes resistant strains to breach the critical bacterial concentration (CBC) above which the immune system can no longer control the resistant bacteria. Unfortunately, the more infectious the phages in a first cocktail, the faster the release from competition, and the faster phage-resistant bacteria will breach this CBC. Indeed, once the infection rate of the phages suffices for controlling sensitive bacteria, a further increase of the infection rate is less important than extending the therapy with additional phages (Fig 6). It is therefore better to provide all phages at the same time [7,36], and use diverse cocktails of functionally different phages [32], to avoid this early expansion of phage-resistant bacterial strains. These findings mimic the lessons learned from antiretroviral therapy to HIV, where one combines several drugs, each requiring different resistance mutations [19], to hit the virus hard and early [18]. The main difference with antiretroviral and antibiotic therapy is that the dosing is less important in phage therapy because phages are rapidly ‘replicating drugs’ (the maximum replication rate of phages, ρ=9.21 h–1, by far exceeds that of their bacterial hosts). This is clinically important because using lower doses would reduce the problem of residual endotoxin in phage preparations. Current recommendations of the Antibacterial Resistance Leadership Group (ARLG) in the United States are to use the highest safe and tolerated dose of bacteriophage products, although it is not clear that clinical outcome always improve with higher doses [2]. For instance, phage therapies in the United States typically use much higher doses that those in Europe, with similar clinical success rates between 55 and 70% [2]. Trials investigating the effect of dosage are ongoing [37], but the lowest doses are typically kept above 107 PFU/ml. The main concern with using lower doses is that the phages may not reach the site of infection (which in our well-mixed model is never an issue). Nevertheless, because lower doses should reduce current problems with residual endotoxin in bacteriophage products, it would be interesting to further investigate this experimentally and/or clinically. To explain the very rapid development of phage-resistance in the patient suffering from the Acinetobacter infection, both in vivo and in vitro [4,7], we required that a fairly large fraction (>1%) of the pretreatment bacteria were resistant to the first cocktail (see Fig 5). This is high compared to other estimates for the frequency of phage-resistant bacteria in bacterial populations [1,10]. One potential explanation for this difference would be that Acinetobacter baumannii has been involved in a long arms race with T4-like myophages, and has evolved multiple distinct defense systems [33]. Mechanisms allowing some Acinetobacter baumannii to rapidly shed their envelope could be an example of such an adaptation, as this confers resistance to many different phages. Not all defense systems will be present in every individual bacterium, but since several are definitely expected to be present in a large bacterial population, it seems best to chose phages that are equipped to overcome those [33]. Although we have tried to use realistic parameter values by (1) adopting parameter values estimated from phage experiments in mice [10], (2) improving the estimated killing rate of neutrophils by refitting the experiments of Drusano et al. [24] (see S1 Fig), (3) estimating the fitness costs and pretreatment levels of resistance by fitting time courses from in vitro experiments [4] (see Fig 8), (4) estimating the clearance rate of phages from human data [4] (see Fig 9), and (5) bounding the maximum replication rate of phages to ρ=9.21 h–1, several parameter estimates remain uncertain. We also tried to develop a novel model of phage therapy in humans by (1) adopting an appropriate model developed for describing phage experiments in mice [10], (2) allowing for several phages groups and many bacterial strains being sensitive, partially resistant, or fully resistant, (3) improving the neutrophil model by making it dependent on a neutrophil source from the bone marrow, but nonetheless the model remains a caricature of the complex reality of a bacterial infection distributed over various organs. The model lacks other anti-bacterial immune responses, like macrophages that can play an important role during phage therapy [38], the complement system, antibodies, and B and T cells, and lacks antibodies elicited by the phages [39]. It therefore remains unclear whether or not our description of the phage therapy of the patient in Schooley et al. [4] is realistic. We prefer to see these simulations as examples of mechanisms that are most likely to have played a role, and we have used these results as an inspiration for further improvement of future phage therapies. To make this model useful for interpreting future phage therapies it would probably be good to also allow for partial resistance (as the current resistance matrix is binary). Importantly, we also need better estimates of several of its crucial parameters, which requires detailed measurements during a phage therapy. For instance, knowing the bacterial resistance mechanisms for each of the cocktails [7], was essential for establishing the actual number of phage groups in our model. It would therefore be very useful to measure the time courses of the densities of bacterial strains having different resistance mechanisms, and the density of the various phages and neutrophils over time. Such longitudinal data before, during and after treatment, would allow the model to better quantify the various processes determining the successfulness of future phage therapies, and help to determine exactly how such a phage therapy manages to clear a rampant bacterial infection. Despite these uncertainties on the parameter values and mechanistic details of our models, our major observation that the outgrowth of phage-resistant bacterial strains is fast because they are released from competition with the dominant pool of sensitive bacteria, is so general that it does not depend on the details of our model. The hitting hard and early [18,36] with cocktails setting a high genetic barrier [19,32], is a recommendation applying to any species that rapidly evolves resistance to a treatment. 4 Methods 4.1 The initial replication rate The initial replication rate of this model can be computed by considering the initial phase of a small population of phages (P≪hP) during a rampant bacterial infection (S=K≫hB), i.e., by simplifying Eq. (12) into dIdt=βhBP−dII. The initial phase of the infection is then described by the linear system, with eigenvalues which has only one positive root defining the initial replication rate Solving for the new saturation constant, we find (See Gadhamsetty et al. [40] who used a similar approach for estimating the infection rate, β, during the acute phase of an HIV-1 infection). Substituting the parameters given in the captions of Figs 1 and 2, and considering the upper bound ρ=9.21 h–1, we find that in mice this replication rate is approached when hB=9.63×106 CFU/g (which is similar to the other saturation constant, hP=1.5×107 PFU/g, in Eq. (9)). 4.2 Neutrophil killing rates Unfortunately, the estimates for the killing rates suffer from several uncertainties. First, Drusano et al. [24] actually fitted their data with a different model having a saturated per capita killing rate, rather than a saturated total killing rate. Second, the data they used was relatively sparse and only consisted of 5 data points, i.e., 5 pairs of an initial bacterial density, B(0), and a corresponding density, B(24), after 24 h (see the circles in S1A Fig). Apparently their fitting procedure suffered from a parameter identifiability problem, because the standard deviations of most parameter estimates are larger than their mean [24]. We have refitted the data with Eq. (15), while reducing the identifiability problem by (1) fixing the carrying capacity to the one estimated previously, and (2) by estimating the growth rate, r, from an experiment where Drusano et al. [24] depleted the neutrophils and observed bacterial densities to increase from B(0)=3×104 to B(24)=4.7×107 CFU/g (see the bullets in S1A Fig). Since the latter is well below the carrying capacity, K = 1010 CFU/g, we assume exponential growth, and solve for r from B(24)=B(0)er24 giving r = 0.306 h–1 (i.e., a doubling time of 2.3 h). Fixing r and K we obtain the estimates κ=0.5 h–1 and hk=1.5×107 CFU/g (see Table 2 and the legend of S1 Fig for additional information). Even the current estimates of κ and hk are not completely reliable, because even after fixing r the estimates of k and hk are corrected negatively. Finally, having somewhat better estimates for the total killing rate κ, we still need to estimate the killing rate per neutrophil, k=κ/N, as the neutrophil density should also be increasing in the Drusano et al. [24] experiments (i.e., κ/Nmaxr, implying that k>r/N(0)=1.1×10−7 g/(h cell). Setting k=1.5×10−7 g/(h cell), which is almost 2-fold larger than the previous estimate, we next constrain the estimate of hk from the fact that the lower CBC, B^N0, should be much smaller than 107 CFU/g (as higher densities grow uncontrolled [24]). Solving hk from B^N0=106=hk(kN(0)/r−1) yields hk=2.86×106 CFU/g, which translates into a maximum killing rate of khk=0.43 CFU/g per neutrophil per hour. For these new parameter estimates the upper CBC becomes B^Nmax≃3.1×107 CFU/g. 4.3 Derivation of the model To derive the infection term mechanistically we separate the time scale of phages binding bacteria from the time scale governing the cellular dynamics. First, consider a group of phages, Pi with i=1,2,…,n, that are using the same type of receptor molecule to enter and infect the bacteria. It is important to sub-divide the phages into such groups, because there will be competition amongst phages binding the same type of receptor, and if the bacteria down-regulate this receptor as a phage-resistance mechanism, it will equally affect all phages in this particular group. Next consider one bacterium that is expressing a total of T copies of this particular receptor molecule. The rate at which the cell will become infected by one of the phages is determined by the number of phages bound to a receptor, and becoming internalized. Using R for the number of unbound receptors, ai for the binding (or absorption) rate of phage i, and ei for the entry rate of phage i, we consider the scheme where Ci with i=1,2,…,n, defines the number of receptors bound by phages of type i, and Pi′ defines the corresponding number of internalized phages (see the cartoon in Fig 7A). The infection rate, F, of the bacterium is assumed to be proportional to the number of internalized phages, i.e., F=β∑inPi′. A cartoon explaining the model of Eq. (33) (Panel A) and Eq. (34) (Panel B). The heavy horizontal line depicts the bacterial cell wall. The black boxes with different indentations are receptors having different binding sites (triangular, rectangular or oval). Panel A depicts one group of phages, Pi, i=1,…,4 (where the type, i, is reflected by the color), all interacting with the same triangular ligand, that can bind receptors with a triangular indentation. We depict a total of T = 4 receptors carrying a triangular binding site. Bound receptors are indicated as a complexes, Ci, free receptors for this group as R, and internalized phases are called Pi′. The force of infection, F, is proportional to the total number of internalized phages (see Eq. (33)). Panel B extends this into three phage groups, with a total of Ti=3 receptors per group. Free phages Pik only bind receptors of group k (see the complementary shapes), forming complexes Cik, that are internalized to become Pik′ phages that all contribute to the rate of infection (see Eq. (34)). Defining the number of unbound receptors by the conservation equation R=T−∑inCi we write for the receptors bound by phages in group i Making the QSSA dCidt=0 for all the n phages binding this receptor, we arrive at where ai′=ai/ei. Summing these expressions over the phage groups yields or Hence where 0≤C¯i≤T. For the internalized phages we write dPi′dt=eiC¯i−d′Pi′, where d′ is their intracellular decay rate. Also making a QSSA dPi′dt=0 we finally arrive at (For a single type of phage this can be written as the saturated infection rate of Eq. (1), i.e., as βP/(1+P/hP), where β=a1T/d′ and hP=e1/a1). Since P¯i′ provides the number of internalized phages of type i, we still need to sum over all n phage types to arrive at the infection rate imposed by all phages binding this receptor, i.e., where β is an infection rate of the bacterium, F is its force of infection, and βi=βai/d′ combines the infection rate, with the binding rate, the entry rate, and the decay rate, and aj′ remains the ratio of the binding rate over the entry rate of phage j (see Fig 7A). Since different phage groups may bind different receptors (see Fig 7B), each expressed with Tk copies, the infection rate of Eq. (33) still needs to be summed over all different groups of phages, to obtain the overall infection rate, i.e., where m is the number of phage groups, nk is the number of different phages in group k, and Pik is density of phage i in group k, with corresponding rates βik and aik′. Because we typically have no information about the receptor molecules the phages used in therapies bind, and since Eq. (34) is complicated and requires two different (typically unknown) parameters for each phage, we decided to simplify the infection term by considering all phages within one group as a single phage, having an infection rate, βi, and saturation constant, ei/ai. Summing over the different groups, each containing a single “super phage” the infection term becomes where n now stands for the number of groups, and we have assumed that all groups have the same saturation constant hP=ei/ai, and all receptors have the same copy number, T=Ti. Thus, although the infection rate of each phage group approaches a maximum βihP, adding additional phage groups allows the total infection rate to go up to an expected value nβ¯hP, where β¯ is the average infection rate. Finally, to prevent the infection terms to become unrealistically fast when large numbers of bacteria are each expressing T receptors, we also allow for saturation in the bacterial densities by rewriting Eq. (35) into This double-saturation function can be derived mechanically for systems where a single type of phages interact with a single type of bacteria [29], as in Eq. (9), but for the high-dimensional system many types of phages and bacteria that we consider here this function remains somewhat phenomenological [30]. The main advantage of the double saturation is that the infection term approaches hBβiPi when T→∞, and hPβiT when Pi→∞. 4.4 Fitting the sensitivity assays The activity of the bacteriophage cocktails was tested by growing isolates of A. baumannii sampled from the patient, in the presence and absence of the phages in vitro (see Fig 3 in Schooley et al. [4]. Assuming that infected bacteria hardly replicate and that resistant bacteria are not infected, the relative growth rate observed in these experiments should reflect the relative fitness of the resistant strains. The density of the bacteria was measured over time for 20 h in “relative respirational units,” and assuming that this measure is proportional to the bacterial density, we fit a simple model for bacterial growth to these data. Finally, making the simplifying assumption that the bacterial growth rate depends linearly on the nutrient concentration, we write where B is the bacterial density, 0 < N < 1 the scaled nutrient concentration, e is the amount of resource required per bacterium, r the maximum replication rate, and dB the death rate. Because the observed respiration initially varied around 50 units even in the absence of bacterial growth, we subtracted the minimum of the respiration observed in each experiment, and fitted B(t) to the background-corrected data (see Fig 8). We fit the data in each experiment simultaneously allowing only the replication rate, r, and the initial condition, B(0), to vary between the bacteria with and without all the phages (see the red and black lines in Fig 8, respectively). Differences in the replication rate should reflect fitness costs, and differences in the initial condition should reflect initial densities of phage-resistant bacterial strains present within the isolates. Fitting the sensitivity assays. Activity of the bacteriophage cocktails on bacterial isolates taken before treatment (Panel A) or after six days of treatment (Panel B). Symbols depict the data copied from Fig 3 in Schooley et al. [4] and lines depict the best fits of Eq. (37) to this data. The parameter estimates are provided in Table 3. Fitting was performed by using the Levenberg-Marquardt algorithm implemented in FME [41]. When we fit the model of Eq. (37) to the data of bacteria growing in the absence of phages we obtain an average replication rate of the various strains weighted by their frequencies (the match between the black symbols and the black lines in Fig 8 show that the model describes the data well). In the presence of phages (red symbols in Fig 8), there is initially hardly any expansion of the bacteria, which provides support for our assumption that in these cultures sensitive bacteria are infected so rapidly that they fail to expand, and hence that the estimated replication rate should reflect the growth rate of strains that are resistant to the phages. The model also describes these data well; see Table 3 for the parameter estimates. The first data set of Schooley et al. [4] involved growing bacteria isolated from the patient before the onset of phage therapy with or without the phages comprising the first cocktail ϕPC (see Fig 3A in their paper, and Fig 8A of this paper). The model fits the data with and without ϕPC phages quite well (see Fig 8A), and provides reasonable parameter estimates (Table 3), i.e., a doubling time of about an hour, 0.43 < r < 0.58 h–1, and a negligible death rate, dB=0.042 h–1. The bacteria resistant to the full cocktail have an estimated replication rate, r = 0.43 h–1, which is about 74% of that of the average bacterial population having r = 0.58 h–1 (i.e., in Table 3 we write ρ=0.43/0.58=0.74). Since this is the relative growth with respect to the B00 ‘wild type’ having a fitness ϕ1=1, we set ϕ2=0.74 for the B01 strain. The initial density of the resistant phages is B(0)=1.01, which is about 21% of that of the total B(0)=4.89. The second data set of Schooley et al. [4] is similar except that it uses bacteria isolated from the patient at day six and the second “Navy” cocktail, ϕIV (see Fig 3F in their paper). These data are reproduced in Fig 8B for the bacteria only (black bullets), and for bacteria in the presence of the full ϕIV cocktail (red). The model also fits these two data sets reasonably well (see Fig 8B), and provides similar parameter estimates (Table 3), i.e., a doubling time of about an hour in the absence of phages r≃0.51 h–1, and a negligible death rate, dB=0.051 h–1. In the presence of phages we estimate a replication rate of r = 0.22 h–1, or a relative growth rate of ϕ2=0.22/0.51=0.44 h–1. Since this is the relative growth with respect to the B01 strain having a fitness ϕ2=0.74, we set ϕ4=0.74×0.44=0.33 for the B11 strain. At day six the density of the resistant phages is B(0)=1.95, which is about 30% of that of the total B(0)=6.56. 4.5 Fitting the bacteriophage pharmacokinetics The decay rate of phages in human serum was measured after the administration of 5×109 PFU of bacteriophage via intravenous injection (see Fig 4 in Schooley et al. [4]). We depict this data in Fig 9 by green circles, and fit them in two ways with the following model where we track the concentration of the phages in the inoculum, PI, and phages in the serum, P. Phages are lost from the inoculum at rate d0, a fraction f appears in the serum, where they are lost at rate dP. Setting PI(0)=5×109 and P(0)=0 we fitted P(t) as predicted by Eq. (38) to the observed plasma levels, and estimated d0 = 16 h–1, dP=4 h–1, and f=5×10−6. This best fit is depicted as the red line in Fig 9. We also fitted the same data with simple exponential decay by setting I(0)=0, and estimated P(0)=2.4×104 and dP=3.5 h–1, which is shown as the blue line in Fig 9. Since these two estimates of dP are similar, and the fit of the exponential decay model is somewhat better, we use the estimate based upon the simplest model, dP=3.5 h–1, throughout (see Table 4). The phage titers measured by Schooley et al. [4]. The symbols depict the serum phage concentration data copied from Fig 4 in Schooley et al. [4], the red line the results of fitting of this data with Eq. (38), and the blue line the result of fitting with an exponential decay model. 4.6 Parameter values 10.1371/journal.pcbi.1014408.t004Table 4Values, interpretations and dimensions of the parameters of the human model.NameValueDimensionMeaningReference r 0.31h–1Bacterial replication rate Table 2 dB 0.01h–1Bacterial death rateEstimated b 100—Burst size[10] K 1010CFU/gCarrying capacity[10] dI 2h–1lysis rate bacteria[27,28] dP 3.5h–1Clearance rate phages Fig 9 k 1.5×10−7 h–1Killing rate by neutrophilsTable 2 & [10] hN 107CFU/gSaturation neutrophil increase[10] hk 2.86×106 CFU/gSaturation neutrophil killing Table 2 hP 1.5×107 PFU/gSaturation infection rate[10] hB 3.56×107 CFU/gSaturation infection rate Eq. (24) dN 0.0833h–1Time scale of neutrophils (dN=0.5 d–1)[15,16] σ 2.25×105 cells h–1 σ=N(0)/dN [10] α 9—Defines a 10-fold increase of neutrophils[10,14] cAi 0h–1Clearance rate by antibody i— a 0h–1Absorption rate— βcrit 2×10−8 g/(PFU h)Minimal infection rateComputed ϕ¯ 0.33−1 —Relative fitnesses Table 3 μ 2.85×10−8 —Mutation rate in Fig 3[10] μ 0.0005−0.05 —Mutation rates in Figs 5–6 Table 3 V¯ 103h–1Phage infusion rateEstimated Values, interpretations and dimensions of the parameters of the human model. Supporting information Fitting the bacterial growth data shown in Fig 1 of Drusano et al. [24]. In Panel A the blue line depicts an example of bacterial growth in neutropenic mice (the bullets depict the observed concentrations at t = 0 and t = 24h). Assuming exponential growth this suggests a growth rate of r = 0.306 h–1 (see the main text). The black solid lines in Panel A depict bacterial growth curves starting at 5 different initial densities (the black circles depict the observations). Using these initial densities, the final densities were fitted with the model of Eq. (15) using κ and hk as free parameters, i.e., for K=8.48×1011 CFU/g and r = 0.306 h–1, we estimate κ=0.5 and hk=1.5×107 CFU/g. These predictions are shown as red lines in Panel A and as symbols in Panel B. The horizontal dotted line in Panel A depicts the CBC of Eq. (15), i.e., hk(κ/r−1), for these estimated parameters (which indeed separates densities that expand from those that contract). (TIFF) A continuation of Panels B, D and F from Fig 3. Panel A reveals that the phages recover following their initial decline, and that the system approaches a steady state with sensitive bacteria (B0), resistant bacteria (B1), and phages (P1) after a dampened oscillation that largely involves the phages and the sensitive bacteria. Panels B and C reveal that the phases go extinct during their initial decline (i.e., their densities drop below the threshold θ corresponding to about one phage per 10 kg of tissue). (TIFF) A continuation of the simulations depicted in Fig 4. The phages go extinct in Panels A and B and persist in Panels D and F. Phages go extinct in Panel A because their density drops below the extinction threshold, θ, when the fully resistant bacteria take over. They go extinct in Panel B because the bacteria are cleared. In Panel E where the bacterial infection is also cleared, the phages persist at a steady state determined by the ratio of infusion and loss (i.e., P¯i=V¯/dP). In Panels C, D and F, where the resistant bacteria breach the upper CBC, the phages largely persist by infecting a minor population of sensitive bacteria (this steady state is again approached via a dampened oscillation). (TIFF) 5 Feb 2026 PCOMPBIOL-D-25-02336 Modeling phage therapy PLOS Computational Biology Dear Dr. De Boer, Thank you for submitting your manuscript to PLOS Computational Biology. After careful consideration, we feel that it has merit but does not fully meet PLOS Computational Biology's publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Apr 07 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at ploscompbiol@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pcompbiol/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: * A letter that responds to each point raised by the editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. This file does not need to include responses to formatting updates and technical items listed in the 'Journal Requirements' section below. * A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. * An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, competing interests statement, or data availability statement, please make these updates within the submission form at the time of resubmission. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. We look forward to receiving your revised manuscript. Kind regards, Roland R Regoes Academic Editor PLOS Computational Biology James Faeder Section Editor PLOS Computational Biology Journal Requirements: 1) We ask that a manuscript source file is provided at Revision. Please upload your manuscript file as a .doc, .docx, .rtf or .tex. If you are providing a .tex file, please upload it under the item type u2018LaTeX Source Fileu2019 and leave your .pdf version as the item type u2018Manuscriptu2019. 2) Please provide an Author Summary. This should appear in your manuscript between the Abstract (if applicable) and the Introduction, and should be 150-200 words long. The aim should be to make your findings accessible to a wide audience that includes both scientists and non-scientists. Sample summaries can be found on our website under Submission Guidelines: https://journals.plos.org/ploscompbiol/s/submission-guidelines#loc-parts-of-a-submission 3) Please upload all main figures as separate Figure files in .tif or .eps format. For more information about how to convert and format your figure files please see our guidelines: https://journals.plos.org/ploscompbiol/s/figures 4) We notice that your supplementary Figures, Tables, and information are included in the manuscript file. Please remove them and upload them with the file type 'Supporting Information'. Please ensure that each Supporting Information file has a legend listed in the manuscript after the references list. 5) Please amend your detailed Financial Disclosure statement. This is published with the article. It must therefore be completed in full sentences and contain the exact wording you wish to be published. - State the initials, alongside each funding source, of each author to receive each grant. For example: "This work was supported by the National Institutes of Health (####### to AM; ###### to CJ) and the National Science Foundation (###### to AM)." - State what role the funders took in the study. If the funders had no role in your study, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.". If you did not receive any funding for this study, please simply state: u201cThe authors received no specific funding for this work.u201d Reviewers' comments: Reviewer's Responses to Questions Comments to the Authors: Please note here if the review is uploaded as an attachment. Reviewer #1: The authors build upon an existing mouse model of phage therapy that incorporates immune system synergy, extending it to a human context. The new model includes the use of phage cocktails and accounts for the potential of bacteria to independently evolve resistance to all cocktail components. The modeling approach and parameterization are thoughtfully designed and presented with meticulous clarity, making the work easy to follow and reproduce. Well-parameterized human models are critical for advancing phage therapy development, and this manuscript represents a valuable and timely contribution. I have mostly minor comments, listed below. Model: Would cocktails still always be advantageous if phages might interfere with each other during infection? L146: the latent period varies widely between phages and environments, how important is this assumption for the model results? L 330: Wouldn’t the argument here be similar to the one later in the manuscript where more infectious phages lead to easier resistance evolution? A higher number of phages from the beginning would lead to a similar effect, meaning that the phage number does matter (arguing for a lower initial phage dose). L 345: While probably outside of this manuscript, it would be nice to also consider would partial cross-resistance in a future version of the model. Antibiotics were included as a potential death rate for bacteria but they could also have an influence on phage numbers and infectivity through bacterial physiology and killing of phage-infected hosts. L421: I am confused as to how the adsorption rate a is related to the infectivity vectors beta_j Figures: Fig.1 It might be a bit confusing t only see the sensitive bacteria appear in panel C. Maybe a red S could be written next to the total bacterial line as well to make clear that they are just not visible. And this could be mentioned in the legend. d) The difference between the three lines is hard to see, especially for the neutrophils, maybe changing the line types could help. Fig.2 it feels like switching a and b would allow a more logical ordering Fig.4 B11 seems to have the same colour as N? Fig.5 It would be helpful to show the administration of cocktails 1 and 2 more clearly in the figure. Discussion: The results could be linked back more to previous findings on phage cocktail efficacy, e.g. https://doi.org/10.1099/mic.0.001110, https://doi.org/10.1128/AEM.00757-12 and phage dosing, e.g. https://doi.org/10.1093/cid/ciad516 Writing: While it is very helpful to have a very detailed description of each model component and parametrization step, I am wondering if it would be helpful to separate them out a bit more into Methods and Results parts to improve the flow. E.g. 2.1 seems less of a result than a detailed methods section and could probably be given as a much shorter version in the Results section and an extended one in the Methods. L 228: dB/dT = 0? L482: something seems weird about this sentence Reviewer #2: This is a very clear contribution that makes a compelling case for several key take-home messages. The authors have carefully justified their modelling assumptions throughout, and gone to great lengths to find and explore realistic parameter values. I particularly enjoyed the straight-forward "what have we learned" summaries on the way through the paper. The contribution is a model for writing up results with a clear eye to what we can (and cannot) reliably infer from mathematical models. Minor points: 1. It would be helpful to the reader to mention that the dN/dt equation will be coming later, when the S-R-P system is introduced (perhaps I just missed this). So around eqns 1-3 since (1) depends on N, or around eqns 10-13. 2. I recognize that the SI is already substantial. Nonetheless I found the paper somewhat overly detailed in the parameter estimate section. Table 1 and the associated text, for example, are of interest to readers who are *very* close to this work, whereas the figures for example will be of quite broad interest. The Discussion and Introduction are nicely aimed for broad appeal. Perhaps a little more of the parameter estimation could be directed to the SI, with a broader readership in mind for the main paper? 3. Clarify that the A_j are constants? 4. Figure 3 should be re-arranged such that panels (a) and (b) are in a column, followed by (c & d), then (e & f). Or re-arrange as (3x2) rather than (2x3). Typos, etc: 1. line 219 "These data" 2. atto-fox: I'm very familiar with this phenomenon, but is it sufficiently well-known for your intended readership? Maybe there's an atto-fox reference you could give for the new reader (eg. grad student who has no idea what this might mean)? 3. line 682, double negative "Not all ... need not" Reviewer #3: Review of Boer et al. “Modeling phage therapy” The manuscript by Boer and colleagues is a highly thorough consideration of the bridge between theoretical population dynamic simulations and real-world mouse and patient data. First, the authors develop an ODE model, based on a previously published model by Roach and colleagues, which accounts for the role of the host response (neutrophils) in the dynamics of a single phage treatment of one bacterial population which can evolve resistance. This adapted model is parametrized and ground-truthed against the same mouse data as the Roach model. Next, the authors attempt to retrospectively explain patient outcomes during phage therapy by further updating their model to account for multiple phages and numerous bacterial states. Finally, the authors interpret their simulations to make suggestions about how to successfully employ phage for therapy. Overall, this manuscript represents a strong and important contribution to the field of phage therapy and more generally to the quantitative modeling of antimicrobial treatment and is a good fit for publication in PLOS Computation Biology. That being said, I do have suggestions that the authors should consider in improving their manuscript prior to acceptance. Major Comments: 1. While the authors use the common variables in the field of phage modeling, they are not defined until Table S2. In contrast the states are defined in the text in the first paragraph of the results. This manuscript has the potential to be of interest to mathematical biologists that do not work with phage; thus, the authors should define the variables prior to using them, most likely with a table similar to S2. 2. The authors intend to employ this model generally to phage therapy, thus they should perform a sensitivity analysis for at least some critical parameters which are likely to vary such as b, µ (outside of the context of the multiphage model), k, and the CBC threshold. This analysis also opens up further interpretation by the authors as to what makes a good phage (e.g., if we see the model is very sensitive to a phage specific parameter, then we should consider selecting phage to optimize for that parameter). 3. A more direct and thorough consideration of the limitations of the application of their model (critically, the things their model does not consider) should be brought up in the discussion. This section should be suited for those in the field who are not as familiar with mathematical and computer-simulation modelling. 4. The authors should also expand the second to last paragraph of the discussion to more thoroughly consider what sort of empirical data would further the field based on the predictions of their modelling. The authors should consider both experimental in vitro and clinical data that would be useful. It would not be inappropriate to suggest specific experiments to either further support or reject the predictions of their model. 5. While the authors plausibly explain the biology (phage defense systems) of why the pretreatment resistance frequency is so high in the Patterson case, how can they explain the biology of the incredibly high mutation rate which they use in Figures 5 and 6 (versus the more reasonable rate used in Figure 3)? Minor Comments: 1. While the manuscript certainly does model phage therapy, the title feels a bit ambitious as if it were the definitive manuscript on modeling phage therapy. A slightly more descriptive title should be considered. 2. On my printed copy of the manuscript, the line thickness varies substantially between figures which makes the later figures (e.g., Figure 5 and Figure S3) difficult to interpret. 3. I’d suggest another term instead of “Guessed” for the assumed parameter values in Table S2. 4. The ReadMe for the Zenodo data is not particularly helpful but for recreating the figures. ********** Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes ********** PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] Figure resubmission: -->While revising your submission, we strongly recommend that you use PLOS’s NAAS tool (https://ngplosjournals.pagemajik.ai/artanalysis) to test your figure files. NAAS can convert your figure files to the TIFF file type and meet basic requirements (such as print size, resolution), or provide you with a report on issues that do not meet our requirements and that NAAS cannot fix.-->--> After uploading your figures to PLOS’s NAAS tool - https://ngplosjournals.pagemajik.ai/artanalysis, NAAS will process the files provided and display the results in the "Uploaded Files" section of the page as the processing is complete. If the uploaded figures meet our requirements (or NAAS is able to fix the files to meet our requirements), the figure will be marked as "fixed" above. If NAAS is unable to fix the files, a red "failed" label will appear above. When NAAS has confirmed that the figure files meet our requirements, please download the file via the download option, and include these NAAS processed figure files when submitting your revised manuscript.--> Reproducibility: To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols 20 May 2026 Submitted filename: new_cover_letter.pdf 4 Jun 2026 Dear Prof. De Boer, We are pleased to inform you that your manuscript 'Towards modeling phage therapy' has been provisionally accepted for publication in PLOS Computational Biology. Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests. Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated. IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript. Should you, your institution's press office or the journal office choose to press release your paper, you will automatically be opted out of early publication. We ask that you notify us now if you or your institution is planning to press release the article. All press must be co-ordinated with PLOS. Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Computational Biology. Best regards, Roland R Regoes Academic Editor PLOS Computational Biology James Faeder Section Editor PLOS Computational Biology *********************************************************** PCOMPBIOL-D-25-02336R1 Towards modeling phage therapy Dear Dr De Boer, I am pleased to inform you that your manuscript has been formally accepted for publication in PLOS Computational Biology. Your manuscript is now with our production department and you will be notified of the publication date in due course. The corresponding author will soon be receiving a typeset proof for review, to ensure errors have not been introduced during production. Please review the PDF proof of your manuscript carefully, as this is the last chance to correct any errors. Please note that major changes, or those which affect the scientific understanding of the work, will likely cause delays to the publication date of your manuscript. Soon after your final files are uploaded, unless you have opted out, the early version of your manuscript will be published online. The date of the early version will be your article's publication date. The final article will be published to the same URL, and all versions of the paper will be accessible to readers. For Research, Software, and Methods articles, you will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing. Thank you again for supporting PLOS Computational Biology and open-access publishing. We are looking forward to publishing your work! With kind regards, Zsofia Freund PLOS Computational Biology | Carlyle House, Carlyle Road, Cambridge CB4 3DN | United Kingdom ploscompbiol@plos.org | Phone +44 (0) 1223-442824 | ploscompbiol.org | @PLOSCompBiol