Understanding the evolution of multiple drug resistance in structured populations
Abstract
The evolution of multidrug resistance (MDR) is a pressing public health concern. Yet many aspects, such as the role played by population structure, remain poorly understood. Here, we argue that studying MDR evolution by focusing upon the dynamical equations for linkage disequilibrium (LD) can greatly simplify the calculations, generate more insight, and provide a unified framework for understanding the role of population structure. We demonstrate how a general epidemiological model of MDR evolution can be recast in terms of the LD equations. These equations reveal how the different forces generating and propagating LD operate in a dynamical setting at both the population and metapopulation levels. We then apply these insights to show how the LD perspective: (i) explains equilibrium patterns of MDR, (ii) provides a simple interpretative framework for transient evolutionary dynamics, and (iii) can be used to assess the consequences of different drug prescription strategies for MDR evolution.
Introduction
Antibiotic resistance is one of the biggest current public health problems, with antibiotic resistant infections responsible for tens of thousands of deaths annually (O’Neill, 2015). Of particular concern is the evolution of multidrug resistant (MDR) pathogens, that is, pathogens resistant to multiple classes of antibiotics. Despite its importance, understanding the evolution of MDR remains an ongoing challenge, as it is typically not captured by our understanding of the evolution of single drug resistance (for which there is a large body of theory; e.g., Blanquart, 2019; Bonhoeffer et al., 1997; Lipsitch et al., 2000; Bergstrom et al., 2004; Austin and Anderson, 1999). For instance, suppose we have two drugs, $A$ and $B$, and that a fraction ${f}_{AB}$ of infections caused by the pathogen of interest are resistant to both drugs. To understand MDR evolution, we need to understand what determines the frequency ${f}_{AB}$. If ${f}_{A}$ and ${f}_{B}$ are the frequency of infections resistant to drugs $A$ and $B$, and $D$ denotes any nonrandom association between resistance to drugs $A$ and $B$, then
If $D=0$, then the evolution of resistance to each drug is independent, and so multiple drugs do not qualitatively alter the evolutionary dynamics of single drug resistance. However, whenever $D\ne 0$, understanding the fitness costs and benefits of resistance to each drug in isolation is insufficient to understand the evolution of MDR, because doing so will not tell us what factors govern the propagation of $D$, which in turn will affect ${f}_{A}$ and ${f}_{B}$. Thus the challenge of understanding MDR evolution can be recast as understanding the dynamics of $D$. The quantity $D$ is referred to as linkage disequilibrium (LD), and it has been extensively studied in population genetics (e.g. Lewontin, 1964; Felsenstein, 1965; Ohta, 1982a; Barton, 1995; Rice, 2004; Slatkin, 2008), particularly as it relates to population structure (Ohta, 1982b; Slatkin, 1975; Li and Nei, 1974; Nei and Li, 1973; Lenormand and Otto, 2000; Martin et al., 2006). However, there has been little attempt to apply these insights to MDR evolution; often the dynamics of doubly resistant infections are neglected to simplify the analysis of single drug resistance (e.g. Bergstrom et al., 2004; Bonhoeffer et al., 1997; Beardmore et al., 2017).
Here, we consider a simple epidemiological model of a primarily asymptomatically carried pathogen (e.g. Staphylococcus spp. or Enterococcus spp.) in a structured host population. We show how this model relates to general dynamical equations for LD (Day and Gandon, 2012), in turn revealing the role of population structure in MDR evolution. We then use these equations to show how analyzing problems from the LD perspective: (i) reveals the evolutionary logic underlying patterns of MDR at equilibrium, which we use to build on a recent paper on MDR evolution (Lehtinen et al., 2019); (ii) provides a framework for understanding transient evolutionary dynamics; and (iii) provides insight on the consequences different drug prescription strategies have on MDR, which we apply to a hospitalcommunity setting.
Results
In what follows we will introduce and analyze a model of MDR evolution. We will highlight the most important aspects here while providing more extensive details in the Materials and methods 'Model derivation'. All notation used is summarized in Table 1.
Consider an asymptomatically carried pathogen in a metapopulation consisting of $N$ host populations in which two drugs $d$ are prescribed, specifically, drug $d=A$ and drug $d=B$. Focus upon population $x$. Let ${S}^{x}$ and ${I}_{ij}^{x}$ denote the density of susceptible hosts and $ij$infections, respectively, at time $t$, where $i$ indicates if the infection is resistant ($i=A$) or not ($i=a$) to drug $A$ and $j$ indicates if the infection is resistant ($j=B$) or not ($j=b$) to drug $B$. Susceptible hosts contract $ij$infections at a percapita rate ${\beta}_{ij}^{x}{I}_{ij}^{x}$, where ${\beta}_{ij}^{x}$ is a rate constant, while $ij$infections are naturally cleared at a percapita rate ${\alpha}_{ij}^{x}$. Hosts are treated with drugs $A$, $B$, or both in combination at percapita rates ${\tau}_{A}^{x}$, ${\tau}_{B}^{x}$, and ${\tau}_{AB}^{x}$, respectively. Treatment is instantaneous and resistance is complete, that is, if the host that receives treatment is infected by a strain sensitive to the drug, the infection is cleared instantaneously, whereas if the host that receives treatment is infected by a strain resistant to the drug, treatment has no effect. Hosts move from population $x$ to $y$ at a percapita rate ${m}^{x\to y}$. Transmission between infected hosts leads to superinfection with probability $\sigma $ in which either strain is equally likely to instantaneously outcompete the other (Nowak and May, 1994; Alizon, 2013). We therefore do not allow for prolonged coinfection (Materials and methods 'Model derivation'). Finally, individual infections acquire allele $i$ through either mutation or recombination (during superinfection) at percapita rates ${\mu}_{i}^{x}$ and ${\rho}_{i}^{x}$, respectively (note that ${\rho}_{i}^{x}$ depends upon infection densities, see Materials and methods Equation (13)).
From these epidemiological assumptions, the change in $ij$infections in population $x$ can be written as the sum of four processes
where ${\mathrm{\U0001d7cf}}_{d}$ is equal to 1 if the $ij$infection is resistant to drug $d$ and 0 otherwise (e.g., if $ij=AB$, then the percapita growth is ${r}^{x}+{s}_{A}^{x}+{s}_{B}^{x}+{s}_{E}^{x}$) and $\varphi {\mu}_{ij}^{x}$ and $\varphi {\rho}_{ij}^{x}$ denote the net change in $ij$infections due to mutation and recombination (Figure 1; Materials and methods Equations (11) and (13)). To faciliate comparison with previous results, we have broken the percapita growth term into four components: the ‘baseline’ percapita growth rate, ${r}^{x}$, the (additive) selection coefficients for resistance to drugs $A$ and $B$, ${s}_{A}^{x}$ and ${s}_{B}^{x}$, and any epistatic interactions, ${s}_{E}^{x}$. These latter terms have the standard interpretation. If ${s}_{A}^{x}>0$ (resp. ${s}_{B}^{x}>0$), then resistance to drug $A$ (resp. $B$) is selected for. If ${s}_{E}^{x}>0$, there is positive epistasis, and the percapita growth rate of doublyresistant infections is greater than would be expected by consideration of the percapita growth rate of singlyresistant infections. Thus although Equation (2) is derived from a specific model, the partitioning is very general and applies to many epidemiological scenarios. We stress that any of the terms ${s}_{d}^{x}$, ${s}_{E}^{x}$, $\varphi {\mu}_{ij}^{x}$, and $\varphi {\rho}_{ij}^{x}$ may themselves depend upon population densities (see Figure 1 for a concrete example). Note that this partitioning is not arbitrary, particularly as it applies to the selection coefficients and epistasis. The additive selection coefficients and epistasis are defined in terms of their effect upon fitness. In continuous time models, fitness is percapita growth. Thus, the selection coefficient for allele $k$ measures the additive contribution of allele $k$ to fitness, while epistasis measures the excess of the fitness of strain $AB$ over its value if fitness were additive across the two loci (e.g. Felsenstein, 1965; Karlin, 1975; Rice, 2004; Kouyos et al., 2009) (see also Box 1). Because epistasis is defined in terms of fitness, how costs of resistance are modeled will typically have implications for whether epistasis occurs or not; for example, multiplicative costs will generate epistasis (Box 1; Materials and methods 'Equilibrium analysis of metapopulation consisting of independent populations'). We will return to this point in the examples.
Box 1.
Costs of resistance, epistasis, and multidrug resistance.
The spread of multidrug resistance (MDR) is driven by selection acting on each drug resistance locus, but also on the linkage disequilibrium (LD), which can be produced by epistasis in fitness. Epistasis measures the interaction between resistance alleles (mutations) at different loci and is defined in terms of the percapita growth rates of different genotypes as: ${s}_{E}^{x}\equiv {r}_{AB}^{x}+{r}_{ab}^{x}{r}_{Ab}^{x}{r}_{aB}^{x}$.
Selection at each locus, e.g. ${s}_{A}^{x}={r}_{Ab}^{x}{r}_{ab}^{x}$, depends on the effects of the mutations on the phenotypic traits of the pathogen. However, nonadditive interactions among these mutations can create epistasis (see inset in Figure 1). To better see how these nonadditive effects can emerge, consider the costs of drug resistance on pathogen transmission. Let ${c}_{{\beta}_{d}}^{x}$ denote the parameter controlling the cost of resistance to drug $d$ in population $x$. Then using the notation of Figure 1.
Transmission rates  Epistasis  

${\beta}_{Ab}^{x}$  ${\beta}_{aB}^{x}$  ${\beta}_{AB}^{x}$  $\mathrm{\Delta}{\beta}_{E}^{x}$  
Additive  ${\beta}_{ab}^{x}{c}_{{\beta}_{A}}^{x}$  ${\beta}_{ab}^{x}{c}_{{\beta}_{B}}^{x}$  ${\beta}_{ab}^{x}{c}_{{\beta}_{A}}^{x}{c}_{{\beta}_{B}}^{x}$  0 
Multiplicative  ${\beta}_{ab}^{x}(1{c}_{{\beta}_{A}}^{x})$  ${\beta}_{ab}^{x}(1{c}_{{\beta}_{B}}^{x})$  ${\beta}_{ab}^{x}(1{c}_{{\beta}_{A}}^{x})(1{c}_{{\beta}_{B}}^{x})$  ${\beta}_{ab}^{x}{c}_{{\beta}_{A}}^{x}{c}_{{\beta}_{B}}^{x}$ 
Hence, only multiplicative costs generate nonadditive interactions between loci on transmission, $\mathrm{\Delta}{\beta}_{E}^{x}$, which leads to epistasis (inset of Figure 1); and, in turn epistasis produces LD which affects MDR evolution.
Of course, the magnitude of drug resistance costs and the interaction between these costs at multiple loci need not be additive nor multiplicative. Subject to appropriate constraints on the choice of costs (e.g. $0\le {\beta}_{ij}^{x}\le {\beta}_{ab}^{x}$), our general framework can account for any pattern of epistasis (see Figure 1). The important point is that since epistasis is a key contributor to multilocus evolution, understanding when it occurs and what is producing it (in this case, assumptions about the cost of resistance) can provide valuable insight into the evolutionary dynamics of MDR.
While system (2) contains all the information necessary to analyze MDR evolution, as currently written it is particularly opaque for providing insight. Therefore, we would like to transform it to a form which brings to the forefront the different factors that promote or impede MDR evolution; the way to do this is by focusing upon the dynamical equations for linkage disequilibrium (LD) (Day and Gandon, 2012; Slatkin, 2008). However, the inclusion of multiple populations means that doing so is not as simple as Equation (1) would suggest since there are different scales at which LD and MDR can be measured. As the scale which is of most interest will depend upon the specifics of the problem, in what follows we will consider MDR evolution at both the population and metapopulationlevel.
Populationlevel multidrug resistance
To understand MDR evolution in a given population, say $x$, we need to understand the dynamics of the frequency of infections resistant to drug $A$ and $B$, ${f}_{A}^{x}$ and ${f}_{B}^{x}$, and the dynamics of population LD, ${D}^{x}$. First, consider the dynamics of ${f}_{A}^{x}$ (mutatis mutandis ${f}_{B}^{x}$). Using Equation (2), it is straightforward to compute
where ${I}^{x}$ is the total density of infections in population $x$ and ${f}_{AB}^{x}={D}^{x}+{f}_{A}^{x}{f}_{B}^{x}$ is the frequency of doublyresistant infections. A related formulation to Equation (3) can be found in Day and Gandon, 2012 (see also Rice, 2004).
Equation (3) is partitioned into recognizable quantities. First, if resistance to drug $A$ is selectively advantageous, ${s}_{A}^{x}>0$, then drug $A$ resistance will increase due to direct selection whose strength is dictated by the genetic variance at the locus, ${f}_{A}^{x}(1{f}_{A}^{x})$ (Fisher, 1930). Second, if doublyresistant infections are overrepresented in the population, ${D}^{x}>0$, and resistance to drug $B$ is selected for, ${s}_{B}^{x}>0$, then drug $A$ resistance will increase due to indirect selection upon resistance to drug $B$. Third, if epistasis is positive, ${s}_{E}^{x}>0$, and there is genetic variance at the locus, drug $A$ resistance will increase due to the disproportionate growth of doublyresistant infections. Fourth, mutation and recombination will increase drug $A$ resistance when there is a mutation or recombination bias toward gain of drug $A$ resistance, ${\mu}_{A}^{x}>{\mu}_{a}^{x}$ or ${\rho}_{A}^{x}>{\rho}_{a}^{x}$, and the frequency of infections sensitive to drug $A$ exceeds the frequency of infections resistant to drug $A$, $1{f}_{A}^{x}>{f}_{A}^{x}$. Finally, migration acts to reduce differences between populations.
It follows that drug $B$ treatment alters the predicted dynamics of resistance to drug $A$ via two main effects: (i) the influence of epistasis and (ii) indirect selection on resistance to drug $B$ mediated through the presence of LD (${D}^{x}\ne 0$). Thus, consider the dynamics of ${D}^{x}$,
where ${s}^{x}={f}_{A}^{x}{s}_{A}^{x}+{f}_{B}^{x}{s}_{B}^{x}+{f}_{AB}^{x}{s}_{E}^{x}$ is the average selection for resistance, ${f}_{ab}^{x}=1{f}_{A}^{x}{f}_{B}^{x}+{f}_{AB}^{x}$ is the frequency of doublysensitive infections, and ${\mu}^{x}$ and ${\rho}^{x}$ are the total percapita rates of mutation and recombination, respectively (e.g. ${\mu}^{x}={\mu}_{a}^{x}+{\mu}_{A}^{x}+{\mu}_{b}^{x}+{\mu}_{B}^{x}$; Materials and methods 'Model derivation').
Equation (4) is partitioned into four key processes. First, excess selection for resistance to drug $A$ (resp. $B$), ${s}_{A}^{x}{s}^{x}$, can cause preexisting LD (${D}^{x}\ne 0$) to increase or decrease. For example, if ${s}_{A}^{x}>{s}^{x}$ and ${D}^{x}>0$ then LD will increase. This is because drug $A$ resistant infections are fitter than the average resistant infection and so will increase in frequency. If ${D}^{x}>0$, it is more likely this increase will occur in doublyresistant infections, thereby increasing ${D}^{x}$. Second, mutation and recombination removes any LD present at a rate proportional to the LD (Rice, 2004; Slatkin, 2008). Third, epistasis generates samesign LD, that is, positive epistasis, ${s}_{E}^{x}>0$, leads to MDR overrepresentation, ${D}^{x}>0$ (Felsenstein, 1965; Lewontin and Kojima, 1960; Lewontin, 1964). Positive epistasis could occur if doubleresistance costs are less than expected (Trindade et al., 2009; MacLean et al., 2010; Hall and MacLean, 2011) or drugs are prescribed in combination (Bretscher et al., 2004; Day and Gandon, 2012).
Migration is the final term of Equation (4) and reveals how the metapopulation structure affects population LD. Like epistasis, migration does not require preexisting LD to operate on LD (Li and Nei, 1974; Slatkin, 1975; Feldman and Christiansen, 1974; Ohta, 1982a; Ohta, 1982b). In particular, LD in population $x$ will be generated whenever the frequencies of resistance to drugs $A$ and $B$ differ between population $x$ and any other connected population, say $y$. If both types of resistance are more common in one population than the other $({f}_{A}^{x}{f}_{A}^{y})({f}_{B}^{x}{f}_{B}^{y})>0$, then migration will generate positive LD in both populations, ${D}^{x}>0$ and ${D}^{y}>0$. If instead drug $A$ resistance is more prevalent in one population, while drug $B$ resistance is more prevalent in the other, migration will generate negative LD in both populations.
Notice the presence of the multiplier ${I}^{y}/{I}^{x}$ in the final term of Equation (4). If the populations have roughly the same density of infections, then this term is unimportant. However, when one population, say $y$, has much fewer total infections than population $x$, ${I}^{y}\ll {I}^{x}$, the term ${I}^{x}/{I}^{y}$ will be very large, whereas ${I}^{y}/{I}^{x}$ will be very small. Consequently, the ability of migration to propagate LD will be greater in population $y$ than $x$, and so all else being equal we would predict the population with a lower density of infections will have a greater magnitude of LD than the population with a higher density of infections.
The next insight shows the importance of also taking into account Equation (3). In particular, if we only inspected the migration term of Equation (4) we might conclude that as the percapita migration rate, ${m}^{y\to x}$, increases, so too will the ability of migration to propagate LD. However, the magnitude of population LD is actually maximized at intermediate migration rates (Figure 2). The reason is because the quantity ${m}^{y\to x}$ has two effects. On the one hand, it directly multiplies the migration term in Equation (4) thereby magnifying migration’s potential role in LD buildup, while on the other hand, it also balances infection frequencies between populations (Equation (3)), which in turn will reduce the magnitude of $({f}_{A}^{x}{f}_{A}^{y})({f}_{B}^{x}{f}_{B}^{y})$ in Equation (4). These conflicting forces mean the magnitude of population LD tends to be maximized when migration is neither too infrequent nor too frequent (Figure 2).
Metapopulationlevel multidrug resistance
Now what happens to LD and MDR evolution at the metapopulationlevel? Here we will use $\overline{X}$ to denote the metapopulation average of quantity ${X}^{x}$, e.g., ${\overline{f}}_{A}$ is the average drug $A$ resistance in the metapopulation (see Materials and methods 'Metapopulation LD and MDR' for further details). Using this notation, then analogously to the population case, metapopulation LD is defined as ${D}_{\text{M}}\equiv {\overline{f}}_{AB}{\overline{f}}_{A}{\overline{f}}_{B}$. A more informative, but mathematically equivalent, description of metapopulation LD, however, is to define it in terms of the population variables as
that is, ${D}_{\text{M}}$ is the sum of the average population LD, $\overline{D}$, and the spatial covariance between the frequencies of resistance to drugs $A$ and $B$. Equation (5) shows that even if there is no population LD, that is, ${D}^{x}=0$ and so $\overline{D}=0$, there may still be metapopulation LD; likewise, there may be population LD, ${D}^{x}\ne 0$, but no metapopulation LD, ${D}_{\text{M}}=0$ (Nei and Li, 1973; Ohta, 1982a; Feldman and Christiansen, 1974).
With this in mind, the change in frequency of infections resistant to drug $A$ (mutatis mutandis drug $B$) can be written
The first four terms in Equation (6) are the metapopulationlevel analogues of the first four terms in Equation (3) and so share the same interpretation. The last two terms, however, arise due to spatial heterogeneity in ‘baseline’ growth and selection and so are the consequence of population structure. As these terms are zero in the absence of spatial heterogeneities, they will be our focus here.
First, spatial heterogeneity arises through differences in the ‘baseline’ percapita growth (i.e. ${r}^{x}\ne {r}^{y}$) coupled with differences in the frequencies of drug $A$ resistant infections (i.e. ${f}_{A}^{x}\ne {f}_{A}^{y}$). This is the spatial covariance between ‘baseline’ percapita growth and the frequency of drug $A$ resistant infections, $\text{cov}(r,{f}_{A})$. In particular, more productive populations (larger ${r}^{x}$) will have a disproportionate effect on the change in drug $A$ resistance. For example, if more productive populations also have a greater frequency of drug $A$ resistance, then heterogeneity increases the population frequency of drug $A$ resistance. Heterogeneity in baseline growth could arise through a variety of mechanisms, such as availability of susceptible hosts, treatment rates differences, or pathogen traits (e.g. transmissibility and duration of carriage).
Second, spatial heterogeneity arises through differences in indirect selection for resistance to drug $B$ (i.e. ${s}_{B}^{x}\ne {s}_{B}^{y}$) coupled with differences in the probability that drug $B$ resistant infections are also doublyresistant (i.e. ${f}_{AB}^{x}/{f}_{B}^{x}\ne {f}_{AB}^{y}/{f}_{B}^{y}$). This is the spatial covariance between selection on resistance to drug $B$ and the conditional probability that a drug $B$ resistant infection is doublyresistant, $\text{cov}({s}_{B},{f}_{AB}/{f}_{B})$. In particular, populations experiencing greater selection for resistance to one drug will have a disproportionate effect on the change in frequency of infections resistant to the other drug, whenever populations differ in frequency of doublyresistant infections. As an example, if populations experiencing stronger selection for drug $B$ resistance also have a greater probability of drug $B$resistant infections being doublyresistant, heterogeneity in indirect selection increases the frequency of drug $A$ resistance in the metapopulation.
Next, the dynamics of metapopulation LD can be written as
where $\text{coskew}(r,{f}_{A},{f}_{B})$ is the spatial coskewness between $r$, ${f}_{A}$, and ${f}_{B}$ and we have assumed population differences in mutation and recombination are negligible (see Materials and methods 'Metapopulation LD and MDR'). The first three terms in Equation (7) are the metapopulation level analogues of the first three terms of Equation (4) and so share the same interpretation. The last two terms, however, arise due to spatial heterogeneity in ‘baseline’ growth and selection and so will be our focus here.
First, spatial heterogeneity arises through spatial differences in the ‘baseline’ percapita growth (i.e. ${r}^{x}\ne {r}^{y}$) coupled with spatial heterogeneities in LD (i.e. ${D}^{x}\ne {D}^{y}$) or resistance frequencies (the coskewness term). The logic of the first term is clear: when population LD differs, more productive populations will disproportionately contribute to metapopulation LD. For the second term, when populations covary in frequency of resistance to drug $A$ and $B$, more productive populations will disproportionately contribute to the covariance, $\text{cov}({f}_{A},{f}_{B})$ and so disproportionately contribute to metapopulation LD (through the second term in Equation (5)).
Second, spatial heterogeneity arises through differences in selection for resistance (${s}_{d}^{x}\ne {s}_{d}^{y}$) coupled with differences in the proportion of drug $d$ resistant infections that are doublyresistant (${f}_{AB}^{x}/{f}_{d}^{x}\ne {f}_{AB}^{y}/{f}_{d}^{y}$). The logic here is that populations experiencing stronger selection for resistance are more likely to see an increase in resistant infections. If this increase occurs disproportionately in doublyresistant infections, then from Equation (1) metapopulation LD will increase, whereas if this increase occurs disproportionately in singlyresistant infections, metapopulation LD will decrease. The magnitude of this effect is scaled by ${\overline{f}}_{d}(1{\overline{f}}_{d})$ since selection cannot operate without genetic variation. In the absence of population LD, then ${f}_{AB}^{x}/{f}_{A}^{x}={f}_{B}^{x}$ and ${f}_{AB}^{x}/{f}_{B}^{x}={f}_{A}^{x}$, and so if populations experiencing stronger selection for resistance to one drug also have a greater frequency of infections resistant to the other drug, metapopulation LD will increase. This could occur if, for example, some populations experience greater treatment rates.
As a final note, observe that in contrast to Equation (4), in Equation (7) the percapita migration rates ${m}^{y\to x}$ are nowhere to be found. The reason for this is intuitive: as migration does not affect the total density of infecteds, nor the resistance status of an infection, it will not change the quantities ${\overline{f}}_{AB}$, ${\overline{f}}_{A}$, or ${\overline{f}}_{B}$, and so cannot change metapopulation LD. As a consequence, migration only affects metapopulation LD indirectly by reducing differences in infection frequency between populations, thereby dampening the magnitude (and hence the effect) of $\text{cov}(r,D)$, $\text{cov}(r,{f}_{d})$, and $\text{cov}({s}_{\mathrm{\ell}},{f}_{AB}/{f}_{d})$ in Equation (7). It follows that, all else being equal, the magnitude of ${D}_{\text{M}}$ is a decreasing function of the percapita migration rate, and so is maximized when migration is infrequent (Figure 2).
Modeling the dynamics of LD: why bother?
To this point, we have focused upon developing the LD perspective to provide a conceptual understanding of MDR evolution in structured populations. However, framing the LD perspective in terms of general quantities has meant this conceptual understanding is somewhat abstract. What we now wish to demonstrate, through the consideration of three scenarios, is how the LD perspective can be used to tackle practical problems. In the first scenario, we show how the LD perspective provides additional insight into a recent paper on the effect of spatial structure on equilibrium patterns of MDR. In the second scenario, we show how the LD perspective allows for an understanding of transient dynamics, and we apply this understanding to patterns of MDR observed in Streptococcus pneumoniae. In the third scenario, we show how the LD perspective generates practical insight into designing drug prescription strategies across populations, with a focus upon a hospitalcommunity setting.
LD perspective explains equilibrium patterns of MDR
Understanding the patterns of MDR in structured populations was first tackled in an important paper by Lehtinen et al., 2019. The paper by Lehtinen et al., 2019 (see also Jacopin et al., 2020) focused upon MDR evolution in a metapopulation consisting of independent host populations (so migration is restricted, ${m}^{x\to y}\approx 0$). For example, each population could represent a different Streptococcus pneumoniae serotype maintained by serotypespecific host immunity (HenriquesNormark and Tuomanen, 2013; Cobey and Lipsitch, 2012; Lehtinen et al., 2017). Lehtinen et al., 2019 found that at equilibrium, population differences could lead to MDR overrepresentation (${D}_{\text{M}}>0$), and that populations with a longer duration of pathogen carriage were more likely to exhibit MDR, a result they attributed to an increased likelihood of antibiotic exposure per carriage episode. Here, we show how employing the LD perspective: (i) reveals the evolutionary logic behind what populations differences can maintain metapopulation LD at equilibrium and (ii) using these insights allows us to build upon the results of Lehtinen et al., 2019 to understand how epidemiological factors other than duration of carriage can play an important role. For simplicity, we will assume that costs are additive (see Box 1), and so there is no epistasis (i.e. ${s}_{E}^{x}=0$), but as this differs from Lehtinen et al., 2019 who use multiplicative costs, we discuss this assumption in more depth in Materials and methods 'Equilibrium analysis of metapopulation consisting of independent populations'.
To maintain metapopulation LD at equilibrium, there needs to be at minimum some mechanism maintaining metapopulation resistance diversity, otherwise ${D}_{\text{M}}=0$. There are variety of ways in which this could occur (Lipsitch et al., 2009; Colijn et al., 2010; Davies et al., 2019; Lehtinen et al., 2017; Jacopin et al., 2020; Krieger et al., 2020), but Lehtinen et al., 2017, Lehtinen et al., 2019 assume it is due to some variation among populations in the conditions favoring resistance evolution. This mechanism maintains diversity at the scale of the metapopulation but leads to the fixation or the extinction of drug resistance locally. Thus ${D}^{x}=0$, and it follows from Equation (5) that ${D}_{\text{M}}=\text{cov}({f}_{A},{f}_{B})$. Therefore, in order for metapopulation LD to exist, ${f}_{A}$ and ${f}_{B}$ must covary across populations. Specifically, whenever ${f}_{A}^{x}$ and ${f}_{B}^{x}$ (or their dynamical equations, Equation 3), are uncorrelated, the metapopulation will be in linkage equilibrium. From Equation (3) we see that if the additive selection coefficients, ${s}_{A}^{x}$ and ${s}_{B}^{x}$, are uncorrelated, then so too are the dynamics of ${f}_{A}^{x}$ and ${f}_{B}^{x}$, and so $\text{cov}({f}_{A},{f}_{B})=0$. Hence only when population differences generate correlations between the selection coefficients will they generate LD.
Using this insight, why are populations with a longer duration of carriage associated with MDR (Lehtinen et al., 2019)? And should we expect associations between MDR and any other population attributes? Our primary focus is whether (and how) the selection coefficients are correlated. Letting $\mathrm{\Delta}{z}_{k}^{x}$ be the contribution of trait $z$ to the additive selection coefficient for resistance to drug $d$ in population $x$ (e.g. $\mathrm{\Delta}{\beta}_{A}^{x}={\beta}_{Ab}^{x}{\beta}_{ab}^{x}$), then it is straightforward to compute (see Materials and methods 'Equilibrium analysis of metapopulation consisting of independent populations'),
where we have used slightly different notation from Lehtinen et al., 2019. Now, consider a scenario in which both the treatment rates and the parameters controlling the (additive) costs of resistance are uncorrelated (i.e. $\mathrm{\Delta}{\beta}_{d}^{x}=\mathrm{\Delta}{\beta}_{d}$, $\mathrm{\Delta}{\alpha}_{d}^{x}=\mathrm{\Delta}{\alpha}_{d}$ and ${\tau}_{d}^{x}={\tau}_{d}$); this is one of the scenarios presented in Figure 4 of Lehtinen et al., 2019, with the key difference that they considered ‘multiplicative’ rather than ‘additive’ costs. From Equation (8), the only remaining source of correlation is susceptible density, ${S}^{x}$, which plays a role whenever there are explicit transmission costs, $\mathrm{\Delta}{\beta}_{d}<0$. Although Equation (8) always holds, in keeping with Lehtinen et al., 2019 if we focus upon the equilibrium case, ${S}^{x}$ will be determined by pathogen traits such as transmission and duration of carriage, such that ‘fitter’ populations (i.e. those in which pathogens are more transmissible or have longer duration of carriage) will more substantially deplete susceptibles. By reducing ${S}^{x}$, ‘fitter’ populations lower the transmission costs for resistance to either drug, and so doubleresistance is more likely to be selectively advantageous, even when treatment rates are uncorrelated. In turn, this overrepresentation of doublyresistant infections will generate metapopulation LD.
Thus, when costs are ‘additive’ (Box 1), although variation in duration of carriage can lead to MDR evolution and LD through its effect upon susceptible density (Figure 3a), it is neither necessary (the same pattern can be produced by variation in transmissibility; Figure 3b) nor sufficient (variation in duration of carriage has no effect without explicit transmission costs, Figure 3c). More broadly, if there are more than two drugs, then provided that there are explicit transmission costs for resistance to each drug, susceptible density will generate a correlation between all the selection coefficients, which in turn will yield the pattern of ‘nestedness’ observed by Lehtinen et al., 2019. What is critical for this effect to be prominent, however, is (i) the existence of population differences in susceptible density, and (ii) the costs of resistance (i.e. $\mathrm{\Delta}{\beta}_{d}$), are large enough so as to ensure a strong correlation amongst selection coefficients.
LD perspective explains transient patterns of MDR
The predictions of Lehtinen et al., 2019 were used to explain the patterns of MDR observed in surveillance data. One of these data sets was a surveillance study that documented both the serotype as well as antibiotic resistance to a number of different drugs in S. pneumoniae infections sampled in Maela, Northern Thailand (Turner et al., 2012; Lehtinen et al., 2019). Although the prediction of positive (metapopulation) LD was met for most drug combinations (Lehtinen et al., 2019), inspection of the data set reveals significant serotype LD (Figure 4). This is notable because, as we have detailed above, at equilibrium the simplest version of the model used in the previous section will result in each serotype being in linkage equilibrium, ${D}^{x}=0$. How can we reconcile these conflicting observations? Although there are various possible explanations (e.g. an additional mechanism capable of maintaining diversity withinserotype), here we focus upon relaxing the assumption that the metapopulation is at equilibrium. That is, we are interested in whether longterm transient dynamics unfolding over months and years could plausibly suggest an alternative explanation for the observed serotype LD.
To do so, consider a metapopulation consisting of independent serotypes, differing in their transmissibility and duration of carriage (as in the model of Lehtinen et al., 2017; Lehtinen et al., 2019). Assume that there is no epistasis and that the additive selection coefficients take the form of Equation (8), where the parameters $\mathrm{\Delta}{\beta}_{d}^{x}$, $\mathrm{\Delta}{\alpha}_{d}^{x}$ and ${\tau}_{d}^{x}$ do not depend upon serotype $x$ (Materials and methods 'Transient dynamics and MDR in streptococcus pneumoniae'). Suppose that initially the metapopulation is treated exclusively with drug $A$ at sufficiently high rates such that resistance to drug $A$ goes to fixation in each serotype, that is, ${\overline{f}}_{A}\to 1$ and ${\overline{f}}_{B}\to 0$, and so ${D}_{\text{M}}=0$. Now suppose at time $t=1000$ months that drug $B$ is ‘discovered’ and subsequently prescribed at a high rate in the metapopulation, while owing to its reduced efficacy, prescription of drug $A$ is reduced. Although the treatment rates do not vary by serotype, serotype differences in transmissibility and duration of carriage mean that the changes to treatment rates will differentially affect serotype density, which in turn will differentially affect the serotypespecific availability of susceptible hosts, ${S}^{x}$. Since the serotypespecific selection coefficients, ${s}_{A}^{x}$ and ${s}_{B}^{x}$, and baseline percapita growth, ${r}^{x}$, directly depend upon ${S}^{x}$, the variation in ${S}^{x}$ introduces heterogeneity in Equation (7), which in turn generates metapopulation LD. Because the selection coefficients are positively correlated (due to the shared dependence upon ${S}^{x}$), the metapopulation LD generated will be positive, that is, ${D}_{\text{M}}>0$ (Figure 5a,d). From Equation (7), once metapopulation LD is generated, it will be amplified by directional selection (first term of Equation 7) which is initially positive since resistance to drug $B$ is favored; this leads to a rapid build up of ${D}_{\text{M}}$ (Figure 5a,d). However, this initial increase in ${D}_{\text{M}}$ is transient; for this particular choice of parameter values, at equilibrium ${D}_{\text{M}}\to 0$. Crucially, however, the changes to ${D}_{\text{M}}$ can unfold over a very long time (here the time units are months), such that surveillance data would detect little change in the metapopulation dynamics and so suggest a population roughly in equilibrium.
Although this scenario can lead to considerable (transient) metapopulation LD, there is still nothing generating serotype LD. Our analysis of Equation (4) revealed two possible (deterministic) mechanisms capable of generating population (serotype) LD. First, migration between populations can lead to metapopulation LD spilling over into population LD. In this example, ‘migration’ between serotypes would correspond to serotype ‘switching’ (Croucher et al., 2015), whereby infections exchange serotypes through recombination. However, this is an unlikely explanation as the rate of serotype switching would have to be unrealistically large for serotype LD to be substantially altered. The second term in Equation (4) capable of generating serotype LD is epistasis, which will generate same sign serotype LD (Felsenstein, 1965; Lewontin, 1964; Lewontin and Kojima, 1960). Indeed, in the model considered, negative epistasis, ${s}_{E}^{x}<0$, generates transient negative serotype LD (Figure 5b,e), while positive epistasis generates transient positive serotype LD (Figure 5c,f; Materials and methods 'Transient dynamics and MDR in streptococcus pneumoniae'). Notably, although negative epistasis produces negative serotype LD (and so $\overline{D}<0$), at the scale of the metapopulation this effect is swamped by the positive covariance in frequency of resistance and so metapopulation LD is positive, ${D}_{\text{M}}>0$. (Figure 5e).
Thus, transient dynamics coupled with epistasis could provide a potential explanation for the significant withinserotype LD observed in S. pneumoniae (Figure 4). More generally, the potential complexity of competing selective pressures associated with multilocus dynamics can lead to prolonged, but transient, polymorphisms and LD, and so surveillance data showing limited temporal change in resistance frequency should be treated cautiously and not assumed to be due to a stable equilibrium.
LD perspective helps identify drug prescription strategies limiting the evolution of MDR
Understanding the evolutionary consequences of different antibiotic prescription strategies across populations can have practical relevance for public health. The populations of interest could correspond to physically distinct groups such as a hospital and its broader community, or different geographical regions (e.g. countries). From a public health perspective, when considering different prescription strategies, a variety of factors must be considered, but in general the goal is to successfully treat as many people as possible, thereby reducing the total burden (Bonhoeffer et al., 1997; Abel zur Wiesch et al., 2014). In this circumstance, the LD in the metapopulation and/or populations can provide important information about the likelihood of treatment success. In particular, for a given population frequency of drug $A$ and drug $B$ resistance, negative LD (MDR underrepresentation) increases the likelihood that if treatment with one drug fails (due to resistance), treatment with the other drug will succeed. On the other hand, positive LD (MDR overrepresentation) increases the likelihood of treatment failure, since a greater proportion of resistant infections are doublyresistant and so cannot be successfully treated with either drug. Equations (4) and (7) show that to generate negative LD, drugs should be deployed in a population specific fashion, that is, drug $A$ should be restricted to some populations and drug $B$ restricted to the remaining populations (see also Lehtinen et al., 2019; Day and Gandon, 2012; Jacopin et al., 2020). Doing so will create a negative covariance in selection, such that resistance to drug $A$ (resp. drug $B$) will be favored in some populations and disfavored in the others. This negative covariance in selection will give rise to negative LD and MDR underrepresentation (Figure 2).
As an application of this principle, consider two populations connected by migration, corresponding to a ‘community’ and a much smaller ‘hospital’. Drug prescription occurs at a fixed (total) rate in each population, while the prescription rate is much higher in the hospital (see Materials and methods 'Contrasting drug prescription strategies in a hospitalcommunity setting'). Consider three antibiotic prescription strategies: (i) drugs can be randomly prescribed to individuals (mixing); (ii) drugs can be prescribed exclusively in combination; or (iii) prescription of drug $A$ and $B$ can be asynchronously rotated between the hospital and community, that is, if the hospital uses drug $A$ then the community uses drug $B$, and vice versa (cycling). As both drugs are prescribed at higher rates in the hospital than the community, both mixing and combination generate a positive covariance in selection across populations, producing positive LD and MDR overrepresentation (see Equation (4); Figure 6). Thus, over the short and longterm, mixing and combination produce similar results: doublyresistant infections are favored, while singlyresistant infections are disfavored (Figure 6). Now consider cycling. When drugs are rotated rapidly between populations, infections in either population are likely to be exposed to both drugs. Because prescription rates are higher in the hospital, this effectively creates a positive covariance in selection (i.e. cycling behaves like mixing) and so when resistance emerges, infections tend to be doublyresistant (MDR overrepresentation). When drugs are rotated less frequently, infections are more likely to be exposed to a single drug, creating a negative covariance in selection across populations. In this circumstance, although single resistance can emerge at lower treatment rates then when rotations are more frequent, the negative LD produced by the negative covariance in selection inhibits the emergence of doubleresistance (MDR underrepresentation; Figure 6).
These results emphasize an important tradeoff: delaying the evolution of MDR (e.g. by decreasing time between rotations) promotes the evolution of single drug resistance, whereas delaying the evolution of single drug resistance promotes the evolution of MDR. This is logical: when we maintain a constant treatment rate per individual, decreasing selection for the ‘generalist’ strategy (MDR) necessarily increases selection for the ‘specialist’ strategy (single drug resistance) (Wilson and Yoshimura, 1994). Thus, cycling can either be the best, or worst, option for single drug resistance (Beardmore et al., 2017), but critically, this has concomitant effects for MDR (see also Figure 6). Indeed, mixing, combination and cycling have been exhaustively compared in the context of single drug resistance (e.g. Bonhoeffer et al., 1997; Lipsitch et al., 2000; Bergstrom et al., 2004; Beardmore and PenaMiller, 2010; Beardmore et al., 2017; Abel zur Wiesch et al., 2014; Tepekule et al., 2017); yet these studies largely ignored the consequences for MDR evolution. Our analysis suggests controlling for single drug resistance will have important consequences for MDR, and so it should not be considered in isolation. More generally, whether it is optimal to either delay single drug resistance or prevent MDR will depend upon what metric is used to evaluate what constitutes a ‘success’ or ‘failure’.
Discussion
The evolution of multidrugresistant pathogens is a pressing health concern and is a topic which is increasingly gaining attention from evolutionary biologists and mathematical modellers alike. However, the typical process in studying the problem of MDR is to introduce a model of the form of (2), and then either proceed to a numerical analysis of these equations or simplify the model further by neglecting the dynamics of double resistant infections (Bergstrom et al., 2004; Bonhoeffer et al., 1997; Beardmore et al., 2017). This is because models of MDR evolution rapidly become intractable, a problem which is particularly acute when incorporating aspects of population structure. Here, we have argued that a more insightful and simplifying approach is the ‘linkage disequilibrium perspective’: after specifying the model of interest, as in (2), it is desirable to transform the model into the form of Equations (3), (4), (6), and (7), which brings to the forefront the role played by linkage disequilibrium for MDR evolution in structured populations. The LD perspective is particularly useful for analyzing and understanding transient evolutionary dynamics (Figure 5), which cannot be understood by, for example, invasion analysis.
Our analysis emphasizes that metapopulation structure alone can generate and maintain LD (and so MDR), even in the absence of epistasis (Nei and Li, 1973; Ohta, 1982a; Li and Nei, 1974; Slatkin, 1975). Since in natural populations metapopulation structure is often hidden (e.g. Rosen et al., 2015), patterns of MDR should not be assumed to be due to epistasis, even if no structure is readily apparent. Moreover, caution must be taken when measuring LD (and MDR) at a particular scale, as doing so can lead to erroneous conclusions: even if the metapopulation is in linkage equilibrium, ${D}_{\text{M}}=0$, the populations need not be, ${D}^{x}\ne 0$, and vice versa (Equation (5)), while in more extreme cases, population and metapopulation LD can be of opposite sign (Figure 5b,e). These are not merely esoteric points; the presence or absence of LD (and MDR), and its source (epistasis or metapopulation structure) is critically important. For example, when MDR is due to metapopulation structure rather than epistasis, prescribing drugs across different populations so as to create a negative covariance in selection can reduce the prevalence of MDR (Figure 6; Day and Gandon, 2012; Jacopin et al., 2020; Lehtinen et al., 2019), while distinguishing between population and metapopulation LD can provide additional insight toward evaluating hypotheses (Figures 4 and 5).
Our analysis assumed that the evolutionary dynamics were deterministic, thus neglecting the influence of stochasticity. However, it is widely appreciated in population genetics that stochasticity can play an important role in multilocus dynamics. For example, LD can be generated through genetic drift (Hill and Robertson, 1966; Barton, 1995; Lenormand and Otto, 2000; Otto and Barton, 2001; Keightley and Otto, 2006; Martin et al., 2006), which in turn can interfere with the strength of selection (Hill and Robertson, 1966; Neher and Shraiman, 2011; Slatkin, 2008). Similarly, the (random) genetic background a rare mutation finds itself upon is critically important for its success (Kouyos et al., 2006; Gillespie, 2000; Neher, 2013), and in finite populations this alone can generate LD. However, little has been done to relate these results to evolutionary epidemiology, or to understand how epidemiological feedbacks can influence their predictions. The little work to date has relied upon complex simulations (e.g. Althaus and Bonhoeffer, 2005; Kouyos et al., 2009), which necessarily sacrifice general insight for specificity. Thus, the role of stochasticity in the evolution of MDR remains an area in which further investigation is warranted.
Understanding the evolution of MDR is a research topic of pressing concern. Here, we have argued that using the linkage disequilibrium perspective leaves us better equipped to determine what factors are responsible for generating MDR, and their generality. Moreover, taking such an approach leads to a more straightforward comparison with existing models and results.
Materials and methods
Here, we provide more comprehensive details on the analysis presented in the main text. We start by deriving the general epidemiological model for the dynamics of the different strains which are characterised by their multilocus genotype (Materials and methods 'Model derivation'). We then convert this model into an equivalent system which tracks the dynamics of allele frequencies at each locus, and the LD at the population level (Materials and methods 'Population LD and MDR'), before considering the set of equations for the dynamics of allele frequencies at each locus and the LD at the metapopulation level (Materials and methods 'Metapopulation LD and MDR').
We conclude by providing a detailed mathematical analysis of the three examples presented in the main text: (1) using the LD perspective to explain equilibrium patterns of MDR (Materials and methods 'Equilibrium analysis of metapopulation consisting of independent populations'); (2) using the LD perspective to explain transient patterns of MDR (Materials and methods 'Transient dynamics and MDR in streptococcus pneumoniae'); and (3) applying the LD perspective to identify drug prescription strategies limiting MDR evolution (Materials and methods 'Contrasting drug prescription strategies in a hospitalcommunity setting').
Model derivation
Request a detailed protocolOur focus is on an asymptomatically carried bacteria species in a metapopulation consisting of $N$ populations. Focus upon an arbitrarily chosen population $x$. Let ${S}^{x}$ and ${I}_{ij}^{x}$ denote the density of susceptible hosts and $ij$infections, respectively, at time $t$, where $i$ indicates if the infection is resistant ($i=A$) or not ($i=a$) to drug $A$ and $j$ indicates if the infection is resistant ($j=B$) or not ($j=b$) to drug $B$. Susceptible hosts contract $ij$infections at a percapita rate ${\beta}_{ij}^{x}{I}_{ij}^{x}$, where ${\beta}_{ij}^{x}$ is a rate constant, while $ij$infections are naturally cleared at a percapita rate ${\alpha}_{ij}^{x}$. Hosts in population $x$ are treated with antibiotics $A$, $B$, or both in combination, at percapita rates ${\tau}_{A}^{x}$, ${\tau}_{B}^{x}$, and ${\tau}_{AB}^{x}$, respectively. Hosts move from population $x$ to population $y$ at a percapita rate ${m}^{x\to y}$.
The resistance profile of an infection changes through two processes. First, there may be de novo mutation, and so let ${\mu}_{i}^{x}$ be the percapita rate at which an infection in population $x$ acquires allele $i$ through mutation. Second, a $ij$infection may be superinfected by a $k\mathrm{\ell}$strain (Day and Gandon, 2012); in this circumstance recombination may occur. Specifically, $k\mathrm{\ell}$strains are transmitted to $ij$infections at rate ${\beta}_{k\mathrm{\ell}}^{x}{I}_{k\mathrm{\ell}}^{x}{I}_{ij}^{x}$, whereupon with probability $\sigma $ superinfection occurs. In the event of superinfection, with probability $1\rho $, recombination does not occur, in which case with equal probability the $ij$infection either remains unchanged or becomes a $k\mathrm{\ell}$infection. With probability $\rho $, recombination does occur, in which case with equal probability the $ij$infection becomes either an $i\mathrm{\ell}$ or $kj$infection. Because our focus is upon the role of population structure, we do not allow for coinfection or withinhost competitive differences based upon resistance profiles (e.g. Davies et al., 2019) but these are straightforward extensions. Moreover, at this stage, we do not make any further specification of the dynamics of uninfected hosts, be they susceptible or recovered, as doing so is not essential for a qualitative understanding of MDR evolution.
Rather than immediately writing down the set of differential equations corresponding to these epidemiological assumptions, we instead group the terms based upon the four biological processes that are occurring. In particular, the change in ${I}_{ij}^{X}$ can be written as the sum of:
The net change due to mutation, denoted $\varphi {\mu}_{ij}^{x}$. As an example, focus upon the change in $Ab$infections in population $x$ due to mutation, $\varphi {\mu}_{Ab}^{x}$. These infections can increase through mutation in one of two ways: (i) $ab$infections acquiring allele $A$ at rate ${\mu}_{A}^{x}{I}_{ab}^{x}$ or (ii) $AB$infections acquiring allele $b$ at rate ${\mu}_{b}^{x}{I}_{AB}^{x}$. On the other hand, ${I}_{Ab}^{x}$ infections are lost due to mutation whenever they (i) acquire allele $a$ at a percapita rate ${\mu}_{a}^{x}$, or (ii) acquire allele $B$ at a percapita rate ${\mu}_{B}^{x}$. Combining this information gives the change in $Ab$infections in population $x$ as
(9) $\varphi {\mu}_{Ab}^{x}={\mu}_{A}^{x}{I}_{ab}^{x}+{\mu}_{b}^{x}{I}_{AB}^{x}({\mu}_{a}^{x}+{\mu}_{B}^{x}){I}_{Ab}^{x},$which is mathematically equivalent to
(10) $\varphi {\mu}_{Ab}^{x}={\mu}_{A}^{x}({I}_{ab}^{x}+{I}_{Ab}^{x})+{\mu}_{b}^{x}({I}_{Ab}^{x}+{I}_{AB}^{x}){\mu}^{x}{I}_{Ab}^{x},$where ${\mu}^{x}\equiv {\mu}_{a}^{x}+{\mu}_{A}^{x}+{\mu}_{b}^{x}+{\mu}_{B}^{x}$ is the percapita mutation rate in population $x$. The only difference between the two formulations is interpretation: Equation (9) shows only mutations which lead to a change in state, whereas Equation (10) shows all possible mutations, even those which do not. This is why the percapita loss term, ${\mu}^{x}$, in (10) can be considered the total percapita mutation rate in population $x$. More generally, we can write $\varphi {\mu}_{ij}^{x}$ as
(11) $\varphi {\mu}_{ij}^{x}\equiv {\mu}_{i}^{x}({I}_{aj}^{x}+{I}_{Aj}^{x})+{\mu}_{j}^{x}({I}_{ib}^{x}+{I}_{iB}^{x}){\mu}^{x}{I}_{ij}^{x}.$The net change due to recombination, denoted $\varphi {\rho}_{ij}^{x}$. Specifically, let ${\rho}_{i}^{x}$ be the percapita rate at which infections gain allele $i$ through recombination. For example, consider ${\rho}_{A}^{x}$. In particular, $ij$infections are challenged by strains carrying allele $A$ at rate $({\beta}_{Ab}^{x}{I}_{Ab}^{x}+{\beta}_{AB}^{x}{I}_{AB}^{x}){I}_{ij}^{x}$. With probability $\sigma $, a superinfection event occurs. Given an superinfection event, with probability $\rho $ recombination happens, in which case with probability $1/2$ the recombinant strain $Aj$ will replace the $ij$infection. Thus
(12) ${\rho}_{A}^{x}=\rho {\displaystyle \frac{\sigma}{2}}({\beta}_{Ab}^{x}{I}_{Ab}^{x}+{\beta}_{AB}^{x}{I}_{AB}^{x}),$and $ij$infections acquire allele $A$ in population $x$ at rate ${\rho}_{A}^{x}{I}_{ij}^{x}$. Therefore, the change in $ij$infections in population $x$ due to recombination is
(13) $\varphi {\rho}_{ij}^{x}\equiv {\rho}_{i}^{x}({I}_{aj}^{x}+{I}_{Aj}^{x})+{\rho}_{j}^{x}({I}_{ib}^{x}+{I}_{iB}^{x}){\rho}^{x}{I}_{ij}^{x}$where ${\rho}^{x}$ is the percapita rate of recombination in population $x$, that is,
${\rho}^{x}\equiv \rho \sigma {\displaystyle \sum _{k\mathrm{\ell}}}{\beta}_{k\mathrm{\ell}}^{x}{I}_{k\mathrm{\ell}}^{x}={\rho}_{a}^{x}+{\rho}_{A}^{x}+{\rho}_{b}^{x}+{\rho}_{B}^{x}.$The net change due to host migration between populations,
(14) ${\displaystyle \sum _{y=1}^{N}}{m}^{x\to y}{I}_{ij}^{x}+{\displaystyle \sum _{y=1}^{N}}{m}^{y\to x}{I}_{ij}^{y}.$The net change due to percapita growth,
${r}_{ij}^{x}\equiv {\beta}_{ij}^{x}{S}^{x}{\alpha}_{ij}^{x}{\mathrm{\U0001d7cf}}_{a}(i){\tau}_{A}^{x}{\mathrm{\U0001d7cf}}_{b}(j){\tau}_{B}^{x}(1{\mathrm{\U0001d7cf}}_{A}(i){\mathrm{\U0001d7cf}}_{B}(j)){\tau}_{AB}^{x}(1\rho ){\displaystyle \frac{\sigma}{2}}{\displaystyle \sum _{k\mathrm{\ell}}}({\beta}_{k\mathrm{\ell}}^{x}{\beta}_{ij}^{x}){I}_{k\mathrm{\ell}}^{x},$where ${\mathrm{\U0001d7cf}}_{i}(j)$ is an indicator variable and is equal to 1 if $i=j$ and 0 otherwise.
With these four processes in hand, the dynamics of infection densities are given by the system of $4N$ differential equations.
Population LD and MDR
Request a detailed protocolIn what follows, we provide more details for the calculations of population LD and MDR. First, we define the following frequencies of infections in population $x$ as
where ${I}^{x}={\sum}_{ij}{I}_{ij}^{x}$ is the total density of infections in population $x$. Using these definitions, the standard measure of linkage disequilibrium in population $x$ is
which is mathematically equivalent to
The three dynamical equations of interest for studying MDR in population $x$ are
System (19) contains a number of quantities that we now define in more detail. First, the (additive) selection coefficient for resistance to drugs $A$ and $B$ in population $x$ are defined as
respectively, while epistasis in population $x$ is ${s}_{E}^{x}={r}_{AB}^{x}+{r}_{ab}^{x}{r}_{Ab}^{x}{r}_{aB}^{x}$. It follows that we can write each of the percapita growth rates, ${r}_{ij}^{x}$, as
This is why ${r}_{ab}^{x}={r}^{x}$ can be thought of as ‘baseline’ percapita growth. We define the average selection for resistance in population $x$ as
Note that the average percapita growth rate in population $x$ is therefore ${r}^{x}+{s}^{x}$, that is, average percapita growth rate is the sum of the ‘baseline’ percapita growth rate and the average selection for resistance.
Metapopulation LD and MDR
Next, consider metapopulation (or total) LD and MDR. First, let ${p}^{x}={I}^{x}/{\sum}_{j=1}^{N}{I}^{j}$ be the fraction of total infections in population $x$. Then the metapopulation quantities equivalent to Equations (16) are
The standard measure of linkage disequilibrium at the level of the metapopulation is
which in terms of the population level variables is
where $\overline{D}$ is the average population LD and $\text{cov}({f}_{A},{f}_{B})$ is the spatial covariance between frequency of resistance to drug $A$ and frequency of resistance to drug $B$.
Using these variables, the three dynamical equations for studying metapopulation MDR are
Note that in the equation $\text{d}{D}_{\text{M}}/\text{d}t$, there are terms involving ${\mathrm{\Lambda}}_{ij}$ which we chose to neglect in Equation (7) given in the main text. These terms are
while
Thus the expression
in the equation $\text{d}{D}_{\text{M}}/\text{d}t$ is the effect upon ${D}_{\text{M}}$ of spatial heterogeneity in mutation and recombination rates (${\mu}_{i}^{x}\ne {\mu}_{i}^{y}$ and/or ${\rho}_{i}^{x}\ne {\rho}_{i}^{y}$) coupled with differences in the proportion of infections with allele $i$ (e.g. $i=A$ or $i=a$) that are resistant to the other drug ($j=B$). In particular, populations in which infections are more likely to acquire resistance through mutation/recombination disproportionately affect metapopulation LD through an increase in doublyresistant infections. However, these terms are likely to be quite small because they require that substantial differences in mutation/recombination rates exist between populations. Since these terms are unlikely to be a significant contributor to the dynamics of ${D}_{\text{M}}$, we ignore them in the main text.
There remains a number of other quantities in system (Equation 26) that we now define in more detail. First, the probability that an infection resistant to drug $d$ is found in population $x$ is
For example, if we apply our variable definitions, it is straightforward to show that
Next, to compute the metapopulationlevel selection coefficients, and mutation/recombination rates, we need to compute the weighted average of the population quantities, where the weights are the probability that an infection of a particular type is in population $x$ (calculated above). Applying this logic, the metapopulationlevel selection coefficients and epistasis are
The average selection for resistance in the metapopulation is
The percapita mutation and recombination rates follow similarly. Recall that ${\mu}_{\mathrm{\ell}}$ and ${\rho}_{\mathrm{\ell}}$ are the percapita rates at which infections gain allele $\mathrm{\ell}$. Thus, for example,
Similar calculations can be made to arrive at ${\overline{\mu}}_{B}$, ${\overline{\mu}}_{b}$, and the various ${\overline{\rho}}_{\mathrm{\ell}}$. The total percapita mutation and recombination rates are
Covariance and coskewness
Request a detailed protocolFinally, we also use a number of covariance terms and a coskewness terms. Let $\mathbb{E}[c]$ denote the expectation of the quantity $c$. Then applying the definition of covariance, we have
Following the same procedure, we can calculate $\text{cov}(r,{f}_{A})$ and $\text{cov}(r,{D}_{\text{M}})$. When the covariance involves quantities that also specifically depend upon particular allele(s), the only difference is that when computing the expectation the probability used is the probability that an allele $\mathrm{\ell}$ is in population $x$. For example,
The covariance terms involving the recombination and mutation rates follow similarly, with the appropriate exchanges of variables. Finally, we have the coskewness term, which can be calculated as
Specific examples
Equilibrium analysis of metapopulation consisting of independent populations
Request a detailed protocolThis is a version of one of the models presented in Lehtinen et al., 2019. The metapopulation consists of $N$ populations. The populations are independent (i.e, there is no migration between populations), and each population is assumed to be of a fixed size of unity, so ${S}^{x}=1{\sum}_{ij}{I}_{ij}^{x}$. Resistance is gained and lost through unbiased mutation occurring at rate μ and there is no recombination. Therefore
Let $\mathrm{\Delta}{z}_{d}^{x}$ and $\mathrm{\Delta}{z}_{E}^{x}$ denote the contribution of parameter $z$ to the additive selection coefficient (for drug $d$resistance) and epistasis, respectively, in population $x$. Specifically,
Then if we let ${r}_{ij}^{x}$ denote the percapita growth term of an $ij$infection in subpopulation $x$ (the first term in brackets in Equation (36)), we can partition this as
where
This notation and formulation differs from that of Lehtinen et al., 2017; Lehtinen et al., 2019 in that they assumed costs were multiplicative, that is,
and
where $0\le {c}_{{\beta}_{\mathrm{\ell}}}^{x}\le 1$ and $0\le {c}_{{\alpha}_{\mathrm{\ell}}}^{x}\le 1$ (note the slightly different notation used for multiplicative costs in 1). There are two consequences of multiplicative costs . First, multiplicative costs produce epistasis. For the model of Lehtinen et al., 2017; Lehtinen et al., 2019:
Thus, in this model, there exists epistasis whenever there is a cost of resistance or drugs are prescribed in combination, ${\tau}_{AB}$. More specifically, transmission costs and combination treatment will produce positive epistasis, while duration of carriage costs will produce negative epistasis. Inclusion of epistasis (through multiplicative costs) is not necessarily a problem, and for epidemiological reasons multiplicative costs may be preferable. Indeed, because epistasis plays a central role in multilocus dynamics, it is valuable to recognize if/when epistasis is occurring. However, our analysis in the main text focused upon how population variation in susceptible densities can create correlations in the selection coefficients, favoring MDR, and so we excluded the possibility of epistasis.
The second consequence of multiplicative costs is that they have implications for the cost of resistance within a population. For the above model, and assuming ${c}_{{\alpha}_{A}}^{x}={c}_{{\alpha}_{A}}$, ${c}_{{\beta}_{A}}^{x}={c}_{{\beta}_{A}}$, the selection coefficient for resistance to drug $A$ can be written
From Equation (43), we see that the costs of resistance depend upon the population’s epidemiological parameters (by assumption). Specifically, ignoring concomitant effects upon ${S}^{x}$, populations that are more transmissible (larger ${\beta}^{x}$) with a shorter duration of carriage (larger ${\alpha}^{x}$) pay higher costs of resistance due to how the cost parameters interact with the epidemiological parameters. For example, if there were no costs to transmission (${c}_{{\beta}_{A}}=1$), then Equation (43) predicts that populations with longer duration of carriage (smaller ${\alpha}^{x}$) are more likely to become resistant, because they pay disproportionately lower costs. Indeed, if the populations represent serotype, than this is an example of epistasis between serotype and resistance, which favors LD between duration of carriage and resistance.
To put this in a biological context, if (for example) the populations correspond to the different capsular serotypes of S. pneumoniae, it is possible that the differences between capsules interact with the mechanism of resistance so as to make resistance more costly for more transmissible capsular serotypes or those capsular serotypes associated with longer duration of carriage (multiplicative costs), but it is also possible that no interaction occurs between the capsule differences and the mechanism of resistance (additive costs) or that resistance is less costly for more transmissible serotypes or for capsular serotypes associated with a shorter duration of carriage.
Irrespective of whether the costs are multiplicative or additive, we would attach the constraints that $0\le {\beta}_{ij}^{x}\le {\beta}_{ab}^{x}$ and $0\le {\alpha}_{ab}^{x}\le {\alpha}_{ij}^{x}$, that is, carriage of one or more resistance alleles will never increase transmissibility or decrease duration of carriage, respectively; otherwise there are no costs. As an example, if we were interested in attaching additive resistance costs to transmission, if we let ${c}_{{\beta}_{d}}^{x}$ and ${c}_{{\beta}_{AB}}^{x}$ be the (additive) transmission cost of resistance to drug $d$ and epistatic transmission cost, respectively, in population $x$, so that
then we could attach the constraints
or simply assume that
There are many possible ways of implementing the constraints. One point to note is that even with additive costs, the choice of constraints can also potentially create epistasis; this could occur in the case of Equation (45).
In Figure 3, we consider three scenarios; whenever possible we choose parameter values to agree with those of Figure 4 in Lehtinen et al., 2019. In each scenario, we assume there are 20 independent populations, that the percapita mutation rate is $\mu ={10}^{10}$, and there is no epistasis, ${s}_{E}^{x}=0$. In subplot 3a, we set ${\beta}_{ab}^{x}=2$, while duration of carriage varies among populations from ${\alpha}_{ab}^{x}=0.25$ to ${\alpha}_{ab}^{x}=1.75$. In subplot 3b we set ${\alpha}_{ab}^{x}=0.5$, while transmission varies among populations from ${\beta}_{ab}^{x}=1$ to ${\beta}_{ab}^{x}=3$. In both subplots 3a and 3b, $\mathrm{\Delta}{\alpha}_{A}^{x}=\mathrm{\Delta}{\alpha}_{B}^{x}=0$, while $\mathrm{\Delta}{\beta}_{A}^{x}=\mathrm{\Delta}{\beta}_{B}^{x}=0.1$. Finally in subplot 3 c, $\mathrm{\Delta}{\beta}_{A}^{x}=\mathrm{\Delta}{\beta}_{B}^{x}=0$, while duration of carriage varies among populations from ${\alpha}_{ab}^{x}=0.25$ to ${\alpha}_{ab}^{x}=1.75$, with $\mathrm{\Delta}{\alpha}_{A}^{x}=\mathrm{\Delta}{\alpha}_{B}^{x}=0.05$.
Transient dynamics and MDR in Streptococcus pneumoniae
Request a detailed protocolHere we use a variant of the model originally proposed by Lehtinen et al., 2017, Lehtinen et al., 2019 in which the populations represent different serotypes. Resistance is gained and lost through unbiased mutation at a percapita rate µ and there is no recombination of resistance loci.
Applying these assumptions and using the notation presented with our model from the main text, this yields
where
is a balancing function intended to mimic the stabilizing effect adaptive host immunity has upon serotype diversity ($\omega $ controls the strength of this effect; see Lehtinen et al., 2017; Lehtinen et al., 2019). The treatment rates in Equation (47) are assumed to be independent of serotype. Note that although we could mechanistically model the susceptible hosts available to each subpopulation, ${S}^{x}$, it is more straightforward and computationally simpler to use the phenomenological model given above in which ${S}^{x}=\nu (I,x)S$. The function $\nu (I,x)$ ensures that ${S}^{x}$ has the two properties we are interested in: (i) there can be variation across populations of available susceptibles, and (ii) this variation is linked to population attributes (e.g. transmissibility and duration of carriage). The primary conclusions of our analysis would hold if a mechanistic model for ${S}^{x}$ were used instead.
If we let ${r}_{ij}^{x}$ denote the percapita growth term of an $ij$infection belonging to serotype $x$ (the first term in brackets in Equation (47)), we can partition this as
where if we use the notation introduced in Equation (37),
For simplicity, we keep total population size constant, and so set $S=1{\sum}_{x=1}^{N}{\sum}_{ij}{I}_{ij}^{x}$.
The simulations in Figure 5 assume the metapopulation is initially treated at percapita rates $({\tau}_{A},{\tau}_{B},{\tau}_{AB})=(0.12,0,0)$, until $t=1000$ when these rates switch to $({\tau}_{A},{\tau}_{B},{\tau}_{AB})=(0.07,0.1,0)$. Other parameters values used are $N=12$, $\omega =3$, $\mathrm{\Delta}{\beta}_{A}^{x}=\mathrm{\Delta}{\beta}_{B}^{x}=0.2$, $\mathrm{\Delta}{\alpha}_{A}^{x}=\mathrm{\Delta}{\alpha}_{B}^{x}=0.05$ and $\mu ={10}^{8}$. Finally, because Streptococcus serotypes differ based upon duration of carriage and transmissibility, and there is evidence of a positive correlation between the two (Weinberger et al., 2009; Zafar et al., 2017), ${\alpha}_{ab}^{x}$ was chosen to assume evenly spaced parameter values from ${\alpha}_{ab}^{x}=0.2$ to ${\alpha}_{ab}^{x}=0.7$, while ${\beta}_{ab}^{x}$ was chosen to assume evenly spaced parameter values from ${\beta}_{ab}^{x}=3.25$ to ${\beta}_{ab}^{x}=3$. Cost epistasis, when it is present, is assumed to solely effect transmissibility (i.e. $\mathrm{\Delta}{\alpha}_{E}^{x}=0$). When there is positive epistasis, $\mathrm{\Delta}{\beta}_{E}^{x}=0.065$, whereas for negative epistasis, $\mathrm{\Delta}{\beta}_{E}^{x}=0.015$.
Contrasting drug prescription strategies in a hospitalcommunity setting
Request a detailed protocolWhen we model the hospital and community, we use Equation (2) and assume the susceptible host density is controlled by
where ${\theta}^{x}$ is the influx of new hosts and $d$ is the background mortality rate.
In the hospital/community model, we assume population $C$ is the ‘community’ and population $H$ is the ‘hospital’. Therefore, we let ${\theta}^{H}=0$, and ${m}^{C\to H}=m/{\sum}_{ij}{I}_{ij}^{C}$ be the rate at which individuals are admitted to the hospital, which is independent of population size. Individuals exit the hospital at a constant rate ${m}^{H\to C}$, so they spend on average $1/{m}^{H\to C}$ time units in hospital (assuming background mortality is low). The specification of the migration rates in this way allows us to ensure the ‘community’ is always much larger than the ‘hospital’. In Figure 6, we assumed that the total prescription rate per population, ${\tau}^{C}$ or ${\tau}^{H}$, was the same for each strategy, and that drugs are prescribed at 15× the rate in the hospital versus the community, that is, ${\tau}^{H}=15{\tau}^{C}$. For ‘mixing’, this means $({\tau}_{A}^{x},{\tau}_{B}^{x},{\tau}_{AB}^{x})=({\tau}^{x}/2,{\tau}^{x}/2,0)$, whereas for ‘combination’, this means $({\tau}_{A}^{x},{\tau}_{B}^{x},{\tau}_{AB}^{x})=(0,0,{\tau}^{x})$. Finally, for ‘cycling’ drug $A$ and $B$ were rotated from hospital to community every 50 time units so that either $({\tau}_{A}^{x},{\tau}_{B}^{x},{\tau}_{AB}^{x})=({\tau}^{x},0,0)$ or $({\tau}_{A}^{x},{\tau}_{B}^{x},{\tau}_{AB}^{x})=(0,{\tau}^{x},0)$ depending on the rotation. Therefore in Figure 6 the period is of length $T=100$. In Figure 6, we numerically integrate the system until $t={10}^{4}$; the final state at $t={10}^{4}$ is then what is plotted for the combination and mixing scenarios, whereas for cycling we plot the average state across the last two rotations (i.e. the average state over the final period, from $t=9,900$ to $t={10}^{4}$). Figure 7 shows how changing the length of time between drug rotations affects the evolution of single and multidrug resistance. Specifically, when drug rotations are frequent (with period of $T=1$), cycling behaves like mixing and so positive LD is produced (Figure 7a). As drug rotations become less frequent (period of $T=24$ and $T=160$), cycling generates a negative covariance in selection, which in turn produces negative LD (Figure 7a). Thus, when drug rotations are more frequent, singledrug resistance is delayed and emerges at higher treatment rates, but the evolution of MDR occurs at lower treatment rates (Figure 7b). When drug rotations are infrequent, singledrug resistance emerges at lower treatment rates, but MDR evolution is delayed, emerging at higher treatment rates (Figure 7b). Parameters used in Figure 6 and Figure 7 were ${\beta}_{ab}^{x}=2$, $\mathrm{\Delta}{\beta}_{A}^{x}=\mathrm{\Delta}{\beta}_{B}^{x}=0.4$, ${\alpha}_{ab}^{x}=0.1$, $\mathrm{\Delta}{\alpha}_{A}^{x}=\mathrm{\Delta}{\alpha}_{B}^{x}=0.02$, $d=0.01$, ${\theta}^{C}=0.2$, ${\theta}^{H}=0$, ${m}^{H\to C}=0.5$, $m=0.2$, $\mu ={10}^{7}$, $\sigma =0$.
Data availability
All data used was from a previously published study (Lehtinen et al. 2019 PLoS Pathogens and Turner et al. 2012 PLoS ONE); this data has been uploaded by those authors to a public repository, we downloaded it from there and have provided the details.

PLoS PathogensS1 File. Resistance profiles and duration of carriage estimates for the Maela dataset.https://doi.org/10.1371/journal.ppat.1007763.s002
References

Studies of antibiotic resistance within the patient, hospitals and the community using simple mathematical modelsPhilosophical Transactions of the Royal Society of London. Series B, Biological Sciences 354:721–738.https://doi.org/10.1098/rstb.1999.0425

Antibiotic Cycling and Antibiotic Mixing: Which One Best Mitigates Antibiotic Resistance?Molecular Biology and Evolution 34:802–817.https://doi.org/10.1093/molbev/msw292

Antibiotic cycling versus mixing: the difficulty of using mathematical models to definitively quantify their relative meritsMathematical Biosciences and Engineering : MBE 7:923–933.https://doi.org/10.3934/mbe.2010.7.923

Evolutionary epidemiology models to predict the dynamics of antibiotic resistanceEvolutionary Applications 12:365–383.https://doi.org/10.1111/eva.12753

Recombination in HIV and the evolution of drug resistance: for better or for worse?BioEssays : News and Reviews in Molecular, Cellular and Developmental Biology 26:180–188.https://doi.org/10.1002/bies.10386

What is the mechanism for persistent coexistence of drugsusceptible and drugresistant strains of Streptococcus pneumoniae?Journal of the Royal Society, Interface 7:905–919.https://doi.org/10.1098/rsif.2009.0400

Withinhost dynamics shape antibiotic resistance in commensal BacteriaNature Ecology & Evolution 3:440–449.https://doi.org/10.1038/s415590180786x

The evolutionary epidemiology of multilocus drug resistanceEvolution; International Journal of Organic Evolution 66:1582–1597.https://doi.org/10.1111/j.15585646.2011.01533.x

The effect of population subdivision on two loci without selectionGenetical Research 24:151–162.https://doi.org/10.1017/S0016672300015184

BookThe Genetical Theory of Natural SelectionClarendon Press.https://doi.org/10.5962/bhl.title.27468

Genetic drift in an infinite population. The pseudohitchhiking modelGenetics 155:909–913.

Epistasis buffers the fitness effects of rifampicin resistance mutations in Pseudomonas aeruginosaEvolution; International Journal of Organic Evolution 65:2370–2379.https://doi.org/10.1111/j.15585646.2011.01302.x

The pneumococcus: epidemiology, microbiology, and pathogenesisCold Spring Harbor Perspectives in Medicine 3:a010215.https://doi.org/10.1101/cshperspect.a010215

The effect of linkage on limits to artificial selectionGenetical Research 8:269–294.https://doi.org/10.1017/S0016672300010156

Factors favouring the evolution of multidrug resistance in BacteriaJournal of the Royal Society Interface 17:20200105–20200114.https://doi.org/10.1098/rsif.2020.0105

General twolocus selection models: some objectives, results and interpretationsTheoretical Population Biology 7:364–398.https://doi.org/10.1016/00405809(75)900258

The evolution of recombination in a heterogeneous environmentGenetics 156:423–438.

Stable linkage disequilibrium without epistasis in subdivided populationsTheoretical Population Biology 6:173–183.https://doi.org/10.1016/00405809(74)900227

The population genetics of antibiotic resistance: integrating molecular mechanisms and treatment contextsNature Reviews. Genetics 11:405–414.https://doi.org/10.1038/nrg2778

Genetic draft, selective interference, and population genetics of rapid adaptationAnnual Review of Ecology, Evolution, and Systematics 44:195–215.https://doi.org/10.1146/annurevecolsys110512135920

Statistical genetics and evolution of quantitative traitsReviews of Modern Physics 83:1283–1300.https://doi.org/10.1103/RevModPhys.83.1283

Superinfection and the evolution of parasite virulenceProceedings. Biological Sciences 255:81–89.https://doi.org/10.1098/rspb.1994.0012

Selection for recombination in small populationsEvolution; International Journal of Organic Evolution 55:1921–1931.https://doi.org/10.1111/j.00143820.2001.tb01310.x

WebsiteTackling a global health crisis: initial stepsThe Reviewon Antimicrobial Resistance. London: Wellcome Trust and Government Ofthe United Kingdom. Accessed September 6, 2018.

BookEvolutionary Theory: Mathematical and Conceptual FoundationsSunderland: Sinauer Associates.

Linkage disequilibriumunderstanding the evolutionary past and mapping the medical futureNature reviews. Genetics 9:477–485.https://doi.org/10.1038/nrg2361

On the coexistence of specialists and generalistsThe American Naturalist 144:692–707.https://doi.org/10.1086/285702
Decision letter

George H PerrySenior and Reviewing Editor; Pennsylvania State University, United States

Sonja LehtinenReviewer; ETH Zurich, Switzerland

Stephen M KisslerReviewer; Harvard TH Chan School of Public Health, United States
Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.
Acceptance summary:
This paper addresses the important question of multidrug resistance evolution, which is of both theoretical and applied interest. The authors' efforts to carefully distinguish population and metapopulation linkage disequilibrium and to develop a framework to rigorously analyze the relationship between the two represent an advance in our understanding of microbial population dynamics.
Decision letter after peer review:
[Editors’ note: the authors submitted for reconsideration following the decision after peer review. What follows is the decision letter after the first round of review.]
Thank you for submitting your work entitled "Understanding the evolution of multiple drug resistance in structured populations" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by a Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Brian J Arnold (Reviewer #2).
Our decision has been reached after consultation between the reviewers. Based on these discussions and the individual reviews below, we regret to inform you that your work will not be considered further for publication in eLife.
We have appended to this message a summary of the two reviewers' comments, including both major and minor concerns. Briefly, there were concerns about the extent of the claims made about the model explaining features of data, the clarity of the model development and notation, and links between the results and previous literature. Please see below for details.
Reviewer 1:
MDR is more common than expected by random chance, especially in particular bacterial species. Understanding how this linkage between resistance elements arises is important for public health but also for generally making sense of genomic data, e.g. classifying genomic patterns as outcomes of specific biological processes. Here, McLeod and Gandon show that spatial heterogeneity across subdivided populations can create LD between resistance elements, even in the absence of epistasis.
The novelty of this manuscript appears to be an extension of Day and Gandon, (2012) to a metapopulation model as well as comparing output of the model to a dataset on pneumococcus, which has sort of served as a model for the ecology and evolution of MDR. One particular result of interest was the observation of simulated transient LD even without epistasis. Though this particular observation is not necessarily new (Martin et al., 2006; Day and Gandon, 2012), casting this metapopulation model in an epidemiological framework appears to be new to my knowledge.
However, the intuition this model provides for D_{tot} > 0 at equilibrium is intriguing: variability in susceptibility density across populations can create covariance in selection coefficients for the two alleles, in turn creating correlations between subpopulation allele frequencies, driving D_{tot} > 0. This insight is extremely general and has nothing to do with the biology of a particular species.
While this isn't a major concern, I wanted to point out what I thought was confusing language that made it take longer for me to understand the results.
1) I found the number of ways to quantify LD a little confusing and done with ambiguous language. D is first used in it's standard form D = P_{ab} – P_{a}*P_{b} (Equation 1), which made it confusing when I then encountered D within the definition of D_{tot} (Equation 5), where D is the weighted average of D = P_{ab} – P_{a}*P_{b} for each of the subpopulations. I think it would be clearer if D perhaps consistently referred to a subpopulation, the average D across subpopulations were instead indicated with a summation in Equation 5.
Also with Equation 5, it took me a while to figure out what cov(p_{a}, p_{b}) actually referred to, as the definition "spatial covariance between resistance to drugs A and B", along with being primed by historical literature, made me think of this covariance as LD (i.e. how frequently alleles are found within the same genome). However, this is referring to covariance of *allele frequencies across subpopulations* and has nothing to do with whether or not they're found in the same genome. More specific/clear language will help readers distinguish between the different covariances studied here: between alleles across genomes (LD) and between allele frequencies across subpopulations (cov(p_{a}, p_{b})).
2) Also, I'm left not knowing the ultimate conclusion of applying the model to the pneumococcal data. On line 226 you claim that transient dynamics and epistasis can explain patterns in pneumococcus, whereas on line 301 you claim that variation in S^{X} across subpopulations can generate LD. Some short conclusion regarding these two claims would be useful.
Citations:
Day & Gandon, (2012). The evolutionary epidemiology of multilocus drug resistance. Evolution. https://doi.org/10.1111/j.15585646.2011.01533.x
Martin, Otto & Lenormand, (2006). Selection for recombination in structured populations. Genetics, 172(1), 593609. https://doi.org/10.1534/genetics.104.039982
1) For the case of additive selection, you provide one particular example of parameters that give rise to transient positive LD (Figure 4), but it would seem that many situations would also give rise to negative LD. Is there a symmetry to the model such that if you were to "integrate" across all possible starting conditions, the expectation would be zero? Or is there an asymmetry that makes transient LD more likely to be positive? I assume the latter if you add epistasis, but it's unclear with additive selection. This would help map intuition from your model to results such as Figure 3, in which it seems there's a skew towards positive LD. Is epistasis absolutely needed to explain this skew?
2) One interesting implication of this work is what to expect when bacterial "populations" are either not defined at all or even incorrectly defined. Since Dtot is very positive under many scenarios (even with negative epistasis! Figure 4), you will observe an excess of positive LD between resistance elements (given these starting conditions in Figure 4, that is).
Reviewer 2:
This is an interesting and important topic and the paper promised to be a new take on it, from the point of view of dynamics of linkage disequilibrium.
My major concerns are (1) clarity of the mathematics, notation and model development and (2) with the claims that the model explains data, without a wide range of simulations or a direct comparison of model to data (for example see below re line 239 onwards).
The model in Equation (2) looks linear and the reader has to see figure 1 to note that there is actually a nonlinear term here (the susceptible * infectious term, and a product of two infectious terms) because the s_{A} etc depend on the prevalence.
The model derivation is cumbersome and far from straightforward; I was unable to directly derive (3) from (2). Instead, for the first three terms of (3) I obtained
dpA/dt = (1pA)(pA sA + pAB (sAB – sA – sB) ) + sB D
instead of the expression in the paper which is
(1pA) (pA sA + sAB pAB) + sB D
This would seem to carry through to the other equations. However my derivation attempts were simply using sA, sAB, sB as given in the text, and putting in the rate of growth of the AB variant (r + sAB) as in (2) for example, without substituting in the expressions from Figure 1 for sA, sB, and sAB. This may work it out (?) However this left me a bit sceptical.
Bottom of page 4: if doublyresistant infections are overrepresented in the population, D^{X} > 0. < > middle page 5 "since D^{X} > 0" Which is it?
Page 10: "…, drug B resistance is often more likely to occur in an infection with a genetic background resistant to drug A."  more likely than what?
Line 239++ : This section makes the claim:
"Thus transient dynamics coupled with epistasis can explain the significant withinserotype LD observed in Streptococcus pneumoniae".
This is a big claim, based on a qualitative paragraph with no direct comparison to data, in which there is an underlying assumption of fixation of resistance to drug A in many serotypes. To my knowledge we often do not see fixation of resistance in Streptococcus pneumoniae or in the Maela data in particular, but rather longterm coexistence. I am not convinced that the dynamics of this model explain the significant LD in the data as claimed.
Equation (2) – there is a note above that rho_{l}^{X} depends on infection densities; the nature of this dependence should be specified.
Page 12: why, if there is no mechanism maintaining withinpopulation diversity, does this necessarily mean D^{X} = 0 (line 282) ?
[Editors’ note: further revisions were suggested prior to acceptance, as described below.]
Thank you for submitting your article "Understanding the evolution of multiple drug resistance in structured populations" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and George Perry as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Sonja Lehtinen (Reviewer #1); Stephen M Kissler (Reviewer #2).
The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.
Essential revisions:
While the reviewers of the previous version of your paper were not available to reconsider the revised manuscript, the reviewers of this version of the paper did take the previous comments and your responses into consideration when composing their own reviews.
Collectively we see much potential in your manuscript, but we request the following points to be addressed at minimum (see below reviews for further details and more suggestions):
1. Expanded explanation of the assumptions behind the LD framework. In particular:
a) Is epistasis necessarily defined in terms of an additive expectation on growth rate? (If this is a standard result in population genetics, a reference to an accessible explanation would be enough).
b) The s coefficients are dependent on variable densities – does this matter for the partitioning and interpretation of equations 3 and 4?
2. If the authors disagree with Reviewer 1's questions about the interpretation that variation in susceptible density explains the effect in example 1 and in Lehtinen et al., 2019, more explanation is needed for the following points:
a) Why variation in clearance rate affects selection for resistance when there is a multiplicative cost on clearance rate, but no cost on transmission rate (for the reinterpretation of Lehtinen specifically)?
b) What is happening in example 2, where the effect is currently explained in terms of serotypespecific susceptible density, but the model does not include serotypespecific susceptibles?
3. For examples 2 and 3, please either explain why additive transmission costs are reasonable (given the concerns highlighted below) or change the modeling of cost.
Reviewer #1 (Recommendations for the authors):
My major scientific recommendations are covered by my review. There are some additional less fundamental points that I am happy to share with the authors if they are interested but won't include here as this review is already quite long.
In terms of presentation, I thought everything was mostly very clear, with the exception of the section on serotype dynamics. Here it would be more helpful to make clear the shift from a structured susceptible population to a shared susceptible population.
Reviewer #2 (Recommendations for the authors):
I was asked to review this manuscript after it had already gone through a round of revisions. My comments reflect both my own initial reading of the manuscript and my assessment of the authors' responses to the previous reviews.
While I did not have access the authors' initial submission, it appears that the authors have sufficiently addressed the previous reviewers' comments.
[Editors’ note: further revisions were suggested prior to acceptance, as described below.]
Thank you for resubmitting your work entitled "Understanding the evolution of multiple drug resistance in structured populations" for further consideration by eLife. Your revised article has been evaluated by George Perry (Senior Editor) and a Reviewing Editor.
The manuscript has been improved but there are some remaining issues that need to be addressed as outlined below in edited comments from Reviewer Sonja Lehtinen, as outlined in points 25 below. In addition, I have reviewed the discussion with the eLife editorial office about the article structure and have considered this question myself, ultimately coming to point 1.
1. Please do create a methods section for this paper. There is no need to remove anything from your current main text. Rather, in the new methods section please first overview the methods even if they are already described in more detail elsewhere in the paper (some repetition is ok, and I agree with the level of detail you have provided in the Results section, given the type of paper) in one or more subheadings, and then import all of the subheadings and the full content of the appendix. Since we don't have page limitations and since this is important to the paper, I think it should be included in the methods section of the main text of the paper, rather than as an appendix.
2. The point the authors are making about density of susceptibles and that this point is not limited to additive costs is understood. However, the section on equilibrium patterns of MDR still should be updated because the difference between additive and multiplicative costs is not explicitly flagged in the main text. (I think the discussion in the Sup Mat is very helpful, but most readers will not read the Supplement). Specifically, when considering equilibrium patterns of MDR, the paragraph starting line 283 is only true for additive costs. Without reading the supplement, this paragraph comes across as a general result. If costs are multiplicative, variation in duration of carriage is both necessary (i.e. variation in transmission rate does not produce LD at equilibrium) and sufficient (explicit transmission costs are not needed). Thus, overall, it is true to say that variation in duration of carriage is not necessary, but not that it is not sufficient. It would be helpful to explicitly make this difference clear somewhere in this section. Please also rephrase the sentence lines 270273 to avoid implying that the costs in Lehtinen et al., are additive.
3. The main text would benefit from a discussion of additive costs. Specifically:
a. The authors make a good point that cost parameters are subject to constraints however they are modelled. The reviewer's point was that these constraints are qualitatively different for additive transmission costs: reasonable additive costs for single resistance can lead to negative transmission rate for dual resistance (as in the example in my review) in the absence of cost epistasis. This suggests that for cost parameters where this is the case, assuming no epistasis cannot be appropriate. Although the specific parameter values the authors use don't give rise to this problem, the authors also present more general results for a model with additive costs and no epistasis, so it would be helpful to highlight this constraint somewhere.
b. The authors' careful consideration of how the specification of costs affects interpretation is indeed very useful and interesting. In addition to this conceptual point, the authors say that they are aiming to highlight mechanisms – e.g. variation in transmission rate – that could plausibly give rise to patterns of MDR. In the case of equilibrium patterns, variation in transmission rate only gives rise to LD when transmission costs are additive. The plausibility of variation in transmission rate as an explanation for equilibrium patterns of MDR therefore depends on the plausibility of additive transmission costs. Therefore, if the authors want to suggest that variation in transmission rate is a plausible explanation for patterns of MDR, it is necessary to include a discussion of whether/how additive transmission costs might arise.
4. Add "per carriage episode" to the summary of the biological interpretation of the duration of carriage effect in Lehtinen et al., 2017/2019 (lines 244245), to make it clear that they were not suggesting that (host) populations with longer durations of pathogen carriage have greater antibiotic exposure. (This point was missed in the original review).
5. One final observation, which is not an essential revision and the authors should only address it if they think it will improve the paper. Points 2 and 3 are both related to the result that a model with a multiplicative transmission cost predicts, at equilibrium, LD between duration of carriage and resistance, but not transmission rate and resistance. Would the authors explain this in terms of variation in the density of susceptibles giving rise to the LD with duration of carriage, but this effect being offset by the epistatic interaction between transmission rate and cost of resistance leading to no LD between transmission rate and resistance? This might be interesting to explicitly discuss in the manuscript.
https://doi.org/10.7554/eLife.65645.sa1Author response
[Editors’ note: the authors resubmitted a revised version of the paper for consideration. What follows is the authors’ response to the first round of review.]
Essential revisions:
Reviewer 1:
MDR is more common than expected by random chance, especially in particular bacterial species. Understanding how this linkage between resistance elements arises is important for public health but also for generally making sense of genomic data, e.g. classifying genomic patterns as outcomes of specific biological processes. Here, McLeod and Gandon show that spatial heterogeneity across subdivided populations can create LD between resistance elements, even in the absence of epistasis.
The novelty of this manuscript appears to be an extension of Day and Gandon, (2012) to a metapopulation model as well as comparing output of the model to a dataset on pneumococcus, which has sort of served as a model for the ecology and evolution of MDR. One particular result of interest was the observation of simulated transient LD even without epistasis. Though this particular observation is not necessarily new (Martin et al., 2006; Day and Gandon, 2012), casting this metapopulation model in an epidemiological framework appears to be new to my knowledge.
However, the intuition this model provides for D^{tot} > 0 at equilibrium is intriguing: variability in susceptibility density across populations can create covariance in selection coefficients for the two alleles, in turn creating correlations between subpopulation allele frequencies, driving D_{tot} > 0. This insight is extremely general and has nothing to do with the biology of a particular species.
Yes, one of the main insights we gain from our analysis is a clarification and a generalisation of the results discussed recently by Lehtinen et al., (2019) about the conditions favouring MDR at equilibrium. We believe our Equation (5) provides a new and useful way to understand the evolution of MDR. We would also like to highlight that the dynamical equations we derive are very general and can be used to understand, and perhaps limit, the spread of MDR. For instance, the final example regarding the impact of cycling antibiotics between a hospital and community is new (previous studies did not study the influence of cycling drugs in a metapopulation on MDR evolution) and opens up new perspectives on the management of MDR resistance.
In our revisions, we extended and rewrote our discussions of the examples and we show more explicitly how equations 3 to 7 can help understand the complex evolutionary dynamics of drug resistance.
While this isn't a major concern, I wanted to point out what I thought was confusing language that made it take longer for me to understand the results.
1) I found the number of ways to quantify LD a little confusing and done with ambiguous language. D is first used in its standard form D = P_{ab} – P_{a}*P_{b} (Equation 1), which made it confusing when I then encountered D within the definition of D_{tot} (Equation 5), where D is the weighted average of D = P_{ab} – P_{a}*P_{b} for each of the subpopulations. I think it would be clearer if D perhaps consistently referred to a subpopulation, the average D across subpopulations were instead indicated with a summation in Equation 5.
Also with Equation 5, it took me a while to figure out what cov(p_{a}, p_{b}) actually referred to, as the definition "spatial covariance between resistance to drugs A and B", along with being primed by historical literature, made me think of this covariance as LD (i.e. how frequently alleles are found within the same genome). However, this is referring to covariance of *allele frequencies across subpopulations* and has nothing to do with whether or not they're found in the same genome. More specific/clear language will help readers distinguish between the different covariances studied here: between alleles across genomes (LD) and between allele frequencies across subpopulations (cov(p_{a}, p_{b})).
We agree and we have rewritten many sections of the manuscript with this comment in mind. Additionally, we changed the notation to try improve the readability. We believe this rewording has significantly increased the clarity of our arguments.
2) Also, I'm left not knowing the ultimate conclusion of applying the model to the pneumococcal data. On line 226 you claim that transient dynamics and epistasis can explain patterns in pneumococcus, whereas on line 301 you claim that variation in S^{X} across subpopulations can generate LD. Some short conclusion regarding these two claims would be useful.
We have rewritten both of these examples in light of the reviewers’ comments, and hope the conclusions now are clearer. Specifically, variation in S^{x} (density of susceptibles) across populations can generate metapopulation LD, and in fact can lead to MDR overrepresentation in the metapopulation. However, this example is focused upon an equilibrium analysis. Consequently, although metapopulation LD exists at equilibrium, because nothing maintains diversity withinpopulations (i.e., withinserotypes), there will be no LD within serotypes at equilibrium. This would be true with or without epistasis; i.e., with epistasis, we would expect at equilibrium that $D^{x} = 0$ even if $D_{M} \not = 0$.
In the next example (the transient + epistasis example), when we consider transient dynamics we are specifically focused upon mechanisms that can generate (transient) LD withinserotype, i.e., $D^{x} \not=0$. As we detail in this example, there are two possibilities: epistasis or migration (serotype recombination or serotype switching), however migration (serotype recombination) is an unlikely explanation because the rate at which this occurs would have to be unrealistically high. Therefore, we focus upon exploring the potential role played by epistasis, and show that it can (transiently) generate LD withinserotype. Of course, since there is nothing maintaining withinhost diversity withinserotype, at equilibrium $D^{x} = 0$ as in the previous example. We believe that the way these examples have been rewritten makes it clearer what we are trying to show.
1) For the case of additive selection, you provide one particular example of parameters that give rise to transient positive LD (Figure 4), but it would seem that many situations would also give rise to negative LD. Is there a symmetry to the model such that if you were to "integrate" across all possible starting conditions, the expectation would be zero?
In this example, the shared dependence of the additive selection coefficients upon S^{x} means that there is a tendency to produce positive (metapopulation) LD, both transiently and at equilibrium. In order for this model to produce negative LD, we would need there to be some (strong) factor generating negative LD to counteract this effect. One possibility would be negative epistasis, but this needs to be very strong, as it needs to overcome the positive covariance between additive selection coefficients induced by the shared dependence on S^{x}.
In terms of the initial conditions, we are starting from a state in which there is no LD, and then changing the selective conditions faced by the metapopulation. This then leads to alleles segregating within the population in response to the new selective pressures, which in turn generates LD.
If we were to instead manipulate the initial conditions (i.e., set up negative LD or positive LD) without changing the selective pressures, this is a slightly different problem and is related to the Hill Robertson effect in which LD can interfere with directional selection. In this case, HR theory predicts negative LD would persist longer in the population than positive LD by weakening directional selection. We now discuss HR effects in more detail in the conclusion.
Or is there an asymmetry that makes transient LD more likely to be positive? I assume the latter if you add epistasis, but it's unclear with additive selection. This would help map intuition from your model to results such as Figure 3, in which it seems there's a skew towards positive LD. Is epistasis absolutely needed to explain this skew?
We detailed the emergence of LD at the metapopulation and at the population scale in the revised version. At the metapopulation scale, the positive LD is due to the heterogeneity in baseline growth rate among serotypes (our equation 7 shows how this leads to D_{M}>0). At the scale of serotypes, we find that, indeed, epistasis is required to generate LD (see figure 5e and 5f).
2) One interesting implication of this work is what to expect when bacterial "populations" are either not defined at all or even incorrectly defined. Since Dtot is very positive under many scenarios (even with negative epistasis! Figure 4), you will observe an excess of positive LD between resistance elements (given these starting conditions in Figure 4, that is).
Yes. We like this comment regarding the fact that we often lack a proper description of the metapopulation structure of microbial communities. We now develop this idea further in the conclusions section. Thanks.
Reviewer 2:
This is an interesting and important topic and the paper promised to be a new take on it, from the point of view of dynamics of linkage disequilibrium.
My concerns are (1) clarity of the mathematics, notation and model development
We reworded several sections and changed the notation to improve the readability of our work. As pointed out above, we extended our discussion of the results in the examples section. We believe this analysis helps to show how our analytical work (equations 3 to 7) can help understand the complex dynamics.
and (2) with the claims that the model explains data, without a wide range of simulations or a direct comparison of model to data (for example see below re line 239 onwards).
We claim that our analysis can help to understand observed patterns. For instance, we provide a general understanding on the conditions promoting MDR. This analysis clarifies the link between heterogeneities among mechanisms of clearance and the evolution of drug resistance.
But we do not try to fit quantitatively observed patterns. This would be feasible in principle (parameter estimation) but, in the present work, our main objective was to present a qualitative understanding of the interplay between selection, migration and recombination.
The model in Equation (2) looks linear and the reader has to see figure 1 to note that there is actually a nonlinear term here (the susceptible * infectious term, and a product of two infectious terms) because the s_{A} etc depend on the prevalence.
We clarified that (line 6667).
The model derivation is cumbersome and far from straightforward; I was unable to directly derive (3) from (2). Instead, for the first three terms of (3) I obtained
dpA/dt = (1pA)(pA sA + pAB (sAB – sA – sB) ) + sB D
instead of the expression in the paper which is
(1pA) (pA sA + sAB pAB) + sB D
This would seem to carry through to the other equations. However my derivation attempts were simply using sA, sAB, sB as given in the text, and putting in the rate of growth of the AB variant (r + sAB) as in (2) for example, without substituting in the expressions from Figure 1 for sA, sB, and sAB. This may work it out (?) However this left me a bit sceptical.
We have changed the notation to avoid this type of confusion. s_{AB} was meant to describe epistasis in fitness (not selection on AB genotypes; which is s_{A} + s_{B} + s_{AB}). We now use the notation s_{E} for epistasis in fitness, and we are very explicit in the main text and in the appendix how we derive the different equations.
Bottom of page 4: if doublyresistant infections are overrepresented in the population, D^{X} > 0. < > middle page 5 "since D^{X} > 0" Which is it?
We have clarified the wording in this paragraph.
Page 10: "…, drug B resistance is often more likely to occur in an infection with a genetic background resistant to drug A."  more likely than what?
We have rewritten this entire example and so this sentence no longer appears.
Line 239++ : This section makes the claim:
"Thus transient dynamics coupled with epistasis can explain the significant withinserotype LD observed in Streptococcus pneumoniae".
This is a big claim, based on a qualitative paragraph with no direct comparison to data, in which there is an underlying assumption of fixation of resistance to drug A in many serotypes. To my knowledge we often do not see fixation of resistance in Streptococcus pneumoniae or in the Maela data in particular, but rather longterm coexistence. I am not convinced that the dynamics of this model explain the significant LD in the data as claimed.
It was not our intent to claim that this simple example can fully explain the data. Our intention was to demonstrate that our dynamical equations help understand what we observe in the simulations. This is key because, at first sight, multilocus models may seem overwhelmingly complex. We show with this example that our analysis helps to understand both the shortterm and longterm dynamics of drug resistance.
Note as well, that for the parameter values we have chosen (taken when possible from the Lehtinen et al., 2019 paper for Streptococcus Pneumonaie), it is straightforward to generate scenarios (as in Figure 5) in which there is a considerable amount of transient polymorphism with or without epistasis in figures 5a, b and c. The point here is that things which look like they are in equilibrium based upon surveillance data need not be.
Moreover, the existing explanation for the Maela data presented in Lehtinen et al. 2019 makes a very clear and testable prediction (as we discuss in the text): there should be no LD withinserotype. Yet this prediction is not met in the empirical data (see Figure 4). This would suggest a number of possibilities, including that either: (1) the populations are in equilibrium, but the coexistence mechanism proposed by Lehtinen et al., is not operating, or (2) the coexistence mechanism proposed by Lehtinen et al. is in effect, but the population is not in equilibrium.
In our paper we focus upon examining whether relaxing the conditions in (2), i.e., that the population is in equilibrium, could plausibly explain the patterns of LD in combination with the coexistence mechanism proposed by Lehtinen et al. We show that it could. We stress that this need not be the only explanation (e.g., condition (1) may be at play), and more work would be needed to definitively answer this question.
However, we recognize that the language we previously used in the manuscript may have been misleading about what we were and were not trying to do, and what our analysis does and does not say. As such, we have carefully gone through the manuscript to ensure that the language is not misleading about any conclusions of our analysis, and we have been more specific about what our objectives were for this example.
Equation (2) – there is a note above that rho_{l}^{X} depends on infection densities; the nature of this dependence should be specified.
We have added a reference on line 5253 indicating the equation in the Sup. Mat. which specifies the full dependencies.
Page 12: why, if there is no mechanism maintaining withinpopulation diversity, does this necessarily mean D^{X} = 0 (line 282) ?
This sentence specifically refers to equilibrium in a deterministic model consisting of independent populations. If nothing maintains withinpopulation diversity, then some combination of alleles will be fixed in population $x$, e.g., $f^{x}_{Ab} = 1$ (using the notation where “f_{ij}” refers to the frequency of genotype ij). Since $D^{x} = f_{AB}^{x} f_{ab}^{x} – f_{Ab}^{x} f_{aB}^{x}$, when $f^{x}_{Ab} = 1 $and so $f_{AB}^{x} = f_{ab}^{x} = f_{aB}^{x} = 0$, then it follows that $D^{x} = 0$.
[Editors’ note: what follows is the authors’ response to the second round of review.]
Essential Revisions:
While the reviewers of the previous version of your paper were not available to reconsider the revised manuscript, the reviewers of this version of the paper did take the previous comments and your responses into consideration when composing their own reviews.
Collectively we see much potential in your manuscript, but we request the following points to be addressed at minimum (see below reviews for further details and more suggestions):
1. Expanded explanation of the assumptions behind the LD framework. In particular:
a) Is epistasis necessarily defined in terms of an additive expectation on growth rate? (If this is a standard result in population genetics, a reference to an accessible explanation would be enough).
Yes, for continuous time models epistasis is defined in terms of additive expectations on growth rate. This is because in continuous time models, fitness is percapita growth rate, r. For discrete time models, epistasis is multiplicative (see e.g., Sup. Info. of Kouyos et al., 2009 Epidemics). We have added a paragraph stating this on lines 7078.
But to see why it is a necessity, denote percapita growth of genotype (strain) jk ∈ {ab,Ab,aB,AB} in population x as ${r}_{\mathrm{\text{jk}}}^{x}$. Consider evolution restricted to locus A when allele b is fixed, and let ${p}_{A}^{x}$ be the frequency of allele A in population x. Then the evolution of allele A is governed by
$\frac{{dp}_{A}^{x}}{dt}=({r}_{\mathrm{\text{Ab}}}^{x}{r}_{\mathrm{\text{ab}}}^{x}){p}_{A}^{x}(1{p}_{A}^{x})$ Therefore allele A will increase in frequency whenever $r}_{\text{Ab}}^{x}>{r}_{\text{ab}}^{x$; thus ${s}_{A}^{x}={r}_{\mathrm{\text{Ab}}}^{x}{r}_{\mathrm{\text{ab}}}^{x}$ is the selection coefficient for allele A. Now epistasis is defined as any interaction between loci affecting fitness. Since ${s}_{A}^{x}$ and ${s}_{B}^{x}$ are the additive contribution to baseline fitness, ${r}_{\mathrm{\text{ab}}}^{x}$ , by strains carrying alleles A or B, the fitness of strain AB is the ‘baseline’ fitness, plus the change due to carrying allele A and B (${s}_{A}^{x}$ and ${s}_{B}^{x}$, respectively), plus any remaining fitness unaccounted for due to interactions between loci, ${s}_{E}^{x}$ (epistasis): ${r}_{\mathrm{\text{AB}}}^{x}={r}_{\mathrm{\text{ab}}}^{x}+{s}_{A}^{x}={s}_{B}^{x}+{s}_{E}^{x}$ Rearranging gives ${s}_{E}^{x}={r}_{\mathrm{\text{AB}}}^{x}{r}_{\mathrm{\text{ab}}}^{x}{s}_{A}^{x}{s}_{B}^{x}={r}_{\mathrm{\text{AB}}}^{x}+{r}_{\mathrm{\text{ab}}}^{x}{r}_{\mathrm{\text{Ab}}}^{x}{r}_{\mathrm{\text{aB}}}^{x}$ Thus epistasis in fitness is defined in terms of the percapita growth rates.
There are many different ways we can model the costs of carrying resistance. All are a priori valid but what we want to stress here is the consequences of the way we model these costs on epistasis and consequently on LD.
For a generic epidemiological model, let β^{x}S^{x} and µ^{x} be the percapita transmission and mortality rate in population x, where S^{x} is the available density of susceptibles. Let ${c}_{{\beta}_{k}}^{x}{S}^{x}$ and ${c}_{{\mu}_{k}}^{x}$ be the decrease in transmission and increase in clearance due to carriage of the resistance allele k. Finally, let ${s}_{E}^{x}$ denote any epistatic interactions.
Then the percapita growth of strain jk is
${r}_{\mathrm{\text{jk}}}^{x}=({\beta}^{x}{1}_{A}\left(j\right){c}_{{\beta}_{A}}^{x}{1}_{B}\left(k\right){c}_{{\beta}_{B}}^{x}){S}^{x}{1}_{a}\left(j\right){\tau}_{B}^{x}{1}_{b}\left(k\right){\tau}_{B}^{x}{\mu}^{x}{1}_{A}\left(j\right){c}_{{\mu}_{A}}^{x}{1}_{B}\left(k\right){c}_{{\mu}_{B}}^{x}+{1}_{A}\left(j\right){1}_{B}\left(k\right){s}_{E}^{x},$ where ${1}_{\U0001d4c1}\left(\U0001d4cf\right)$ is 1 if $1=\U0001d4cf$ and zero otherwise. Therefore ${s}_{k}^{x}={\tau}_{k}^{x}{c}_{{\beta}_{k}}^{x}{S}^{x}{c}_{{\mu}_{k}}^{x},k\in \{A,B\}$ For example, in Lehtinen et al., they de_{ne} percapita growth of strain k in population
i, resistant to the set of antibiotics R_{k} and sensitive to the set S_{k} as
${r}_{k}^{x}={\beta}^{x}\prod _{{\mathrm{\text{j\u03f5R}}}_{k}}^{}{c}_{{\beta}_{j}}{S}^{x}{\mu}^{x}\prod _{{\mathrm{\text{j\u03f5R}}}_{k}}^{}{c}_{{\mu}_{j}}\sum _{{\mathrm{\text{j\u03f5S}}}_{k}}^{}{\tau}_{j}^{x}$ We have slightly modified their notation here, but the gist is the same; we have also ignored the possibility of combination treatment (but see Sup. Mat. 4.1 for an example in which it is included). Therefore in the case of two antibiotics, A and B,
we have ${c}_{{\beta}_{k}}^{x}=(1{c}_{{\beta}_{k}}){S}^{x}$ and ${c}_{{\mu}_{k}}^{x}=({c}_{{\mu}_{k}}1)$ and so
${s}_{k}^{x}={\tau}_{k}^{x}(1{c}_{{\beta}_{k}}){\beta}^{x}{S}^{x}({c}_{{\mu}_{k}}1){\mu}^{x},\phantom{\rule{0.222em}{0ex}}k\in \{A,B\}$ and ${s}_{E}^{x}=(1{c}_{{\beta}_{A}})(1{c}_{{\beta}_{B}}){\beta}^{x}{S}^{x}({c}_{{\mu}_{A}}1)({c}_{{\mu}_{B}}1){\mu}^{x}$ Thus there exists epistasis in fitness whenever there is a cost of resistance. More specifically, transmission costs will produce positive epistasis, while duration of carriage costs will produce negative epistasis. We have added further clarification on lines 716738.
b) The s coefficients are dependent on variable densities – does this matter for the partitioning and interpretation of equations 3 and 4?
The partitioning and interpretation of equations 3 and 4 are unaffected by the s coefficients depending upon variable densities; the s coefficients have to be partitioned as they are from consideration of the single locus dynamics (see equation (1)); and this then forces epistasis to be what it is.
Yet, evolutionary dynamics can be more complicated in evolutionary epidemiology than in classic population genetic models as epidemiological feedbacks can change the sign and magnitude of the selective coefficients and epistasis. Although this creates richer dynamics, the fundamental population genetic principles still apply, and provide a useful interpretative framework.
2. If the authors disagree with Reviewer 1's questions about the interpretation that variation in susceptible density explains the effect in example 1 and in Lehtinen et al. 2019, more explanation is needed for the following points:
a) Why variation in clearance rate affects selection for resistance when there is a multiplicative cost on clearance rate, but no cost on transmission rate (for the reinterpretation of Lehtinen specifically)?
Consider the evolution of resistance to drug A in a metapopulation in which allele b is fixed. From (1), resistance is selected for in population x whenever the selection coefficient is positive, ${s}_{A}^{x}>0$. So resistance is selected for whenever the benefits $\left({\tau}_{A}^{x}\right)$ exceed the costs $({c}_{{\beta}_{A}}^{x}{S}^{x}+{c}_{{\mu}_{A}}^{x})$. This is true when the system has reached an endemic equilibrium but also during transient dynamics.
Now how is variation between the populations generated? This could occur a number of ways; but in all cases, the key is that the selection coefficient, ${s}_{A}^{x},$ differs between populations. For example, variation in treatment rate or in the parameters controlling the costs are sufficient. However, in keeping with Lehtinen et al., 2017, 2019, we assume that ${\tau}_{A}^{x}={\tau}_{A}$ and that the cost parameters are population independent.
Then there are two ways in which variation can be produced, depending upon if the cost parameters are additive or are multiplicative (i.e., do they interact or not with the epidemiological parameters?)
i) Additive costs $({c}_{{\beta}_{A}}^{x}={c}_{{\beta}_{A}},{c}_{{\mu}_{A}}^{x}={c}_{{\mu}_{A}})$ : here costs are ‘independent’ of the population, except insofar as the transmission costs also depend upon density of susceptibles.
From ${s}_{A}^{x}$, variation in resistance (due to selection) across populations is only possible if both c_{βA} > 0 and the susceptibles, S^{x}, show variation.
This was the example we focused upon in the manuscript; the prediction here is that any factor negatively correlated with S^{x} will lower the costs of resistance and so select for resistance. For example, populations with a longer duration of carriage (smaller µ^{x} or longer duration of carriage), or higher transmissibility (larger β^{x}), will more substantially deplete available susceptibles, S^{x}, lowering the costs of resistance, ${c}_{{\beta}_{A}}{S}^{x}+{c}_{{\mu}_{A}},$ and so select for resistance.
ii) Multiplicative costs: here the cost parameters multiplicatively interact with the epidemiological parameters. This was true in the model of Lehtinen et al., 2017, 2019 (detailed above); for this model, by assumption (due to the multiplicative interaction), populations with a longer duration of carriage (smaller µ^{x}) pay disproportionately lower costs to acquire resistance than populations with shorter duration of carriage (larger µ^{x}), and so are more likely to become associated with resistance.
Another way of looking at this is if we suppose the populations correspond to serotype, then by specifying an interaction between ${c}_{{\mu}_{A}}$ and µ^{x}, we are creating epistasis between population (serotype) x and resistance. Epistasis produces LD (in this case between serotype and resistance).
Therefore there are two distinct mechanisms generating variation: through the density of susceptibles, or by assuming multiplicative interactions between the cost parameters and the epidemiological parameters. We focused upon additive costs, because we wanted to emphasize the correlation between selection coefficients due to susceptible densities, but this does not mean that multiplicative costs are not valid.
For example, in the case of S. pneumoniae, there may be a physiological reason why the mechanism of resistance interacts with capsule, such that certain capsule types reduce the efficacy of antibiotic resistance over others, or make resistance more costly. Since different capsule types are associated with different durations of carriage, it is possible that such an interaction might mean that capsules with shorter duration of carriage amplify the costs of resistance.
In the Sup. Mat. on lines 716738 we have clarified this issue; we have also modified the language in the main text (lines 247250, lines 290292) when we are discussing the Lehtinen et al., example to make clear that we are primarily focused upon correlations in variation mediated through susceptibles.
b) What is happening in example 2, where the effect is currently explained in terms of serotypespecific susceptible density, but the model does not include serotypespecific susceptibles?
The pool of available susceptibles depends upon serotype due to the stabilizing function intended to mimic adaptive immunity (this is the function from Lehtinen et al., 2017). So the ν(I,x) generates the fraction of susceptibles that are available to serotype x, i.e., S^{x} = ν(I,x)S. Thus there is variation in available susceptibles between serotypes and so there is serotypespecific susceptible density.
It would be possible to mechanistically model each of the populations of susceptibles, however, this would be computationally difficult and not add anything, as the stabilizing function ensures that there is: (1) the potential for serotypespecific differences in available susceptibles and (2) the variation in available susceptibles is linked to attributes of the serotypes (specifically, fitness).
In the Sup. Mat. on lines 756763 we have added text which addresses this issue.
3. For examples 2 and 3, please either explain why additive transmission costs are reasonable (given the concerns highlighted below) or change the modeling of cost.
‘Additive’ transmission costs are always reasonable, given suitable constraints (e.g., ensuring transmission cannot be negative, as in the example provided by the reviewer); but the requirement for suitable constraints is also true for ‘multiplicative’ costs (e.g., the constraint 0 ≤ c_{β} ≤ 1). In the analysis we conducted in examples 2 and 3, appropriate constraints were attached. The key distinction is the costs have different evolutionary implications (as reviewer 1 correctly pointed out).
More generally, we have reframed the relevant sections so it does not seem as if we are advocating for either additive or multiplicative costs; our intent was to point out that they have different evolutionary interpretations. For example, one consequence of multiplicative costs is that they tend to produce epistasis.
We also have amended the text to stress that the inclusion of epistasis in a model is not ‘problematic’ (as our wording incorrectly suggested in the Sup. Mat.); in fact, in many circumstances it is necessary. Our point was that awareness of the presence/absence of epistasis can guide the interpretation.
[Editors’ note: further revisions were suggested prior to acceptance, as described below.]
The manuscript has been improved but there are some remaining issues that need to be addressed as outlined below in edited comments from Reviewer Sonja Lehtinen, as outlined in points 25 below. In addition, I have reviewed the discussion with the eLife editorial office about the article structure and have considered this question myself, ultimately coming to point 1.
1. Please do create a methods section for this paper. There is no need to remove anything from your current main text. Rather, in the new methods section please first overview the methods even if they are already described in more detail elsewhere in the paper (some repetition is ok, and I agree with the level of detail you have provided in the Results section, given the type of paper) in one or more subheadings, and then import all of the subheadings and the full content of the appendix. Since we don't have page limitations and since this is important to the paper, I think it should be included in the methods section of the main text of the paper, rather than as an appendix.
We have now created a methods section that is the appendix. We have added a paragraph on lines 589600 providing an overview of the contents of the methods. In the main text, we now have changed all the references to refer to the relevant sections in the methods.
2. The point the authors are making about density of susceptibles and that this point is not limited to additive costs is understood. However, the section on equilibrium patterns of MDR still should be updated because the difference between additive and multiplicative costs is not explicitly flagged in the main text. (I think the discussion in the Sup Mat is very helpful, but most readers will not read the Supplement). Specifically, when considering equilibrium patterns of MDR, the paragraph starting line 283 is only true for additive costs. Without reading the supplement, this paragraph comes across as a general result. If costs are multiplicative, variation in duration of carriage is both necessary (i.e. variation in transmission rate does not produce LD at equilibrium) and sufficient (explicit transmission costs are not needed). Thus, overall, it is true to say that variation in duration of carriage is not necessary, but not that it is not sufficient. It would be helpful to explicitly make this difference clear somewhere in this section. Please also rephrase the sentence lines 270273 to avoid implying that the costs in Lehtinen et al., are additive.
We have made a number of changes to address these points. First, we have added a box to the main text that discusses the costs of resistance (specifically, additive vs multiplicative) and their consequences for epistasis on pg. 5.
Next, we have made a number of changes to the wording to clarify the points the reviewer is concerned about. Specifically:
– On line 251, we identify that we are using additive costs
– On line 252 we state that this differs from Lehtinen et al., 2019 who use multiplicative costs
– On line 276277 we identify again that our result here differs from Lehtinen et al., 2019 based upon the choice of costs
– On line 287 we state that our results are for additive costs
– In the caption for figure 3 we identify that we are dealing with additive costs (and no epistasis)
3. The main text would benefit from a discussion of additive costs. Specifically:
a. The authors make a good point that cost parameters are subject to constraints however they are modelled. The reviewer's point was that these constraints are qualitatively different for additive transmission costs: reasonable additive costs for single resistance can lead to negative transmission rate for dual resistance (as in the example in my review) in the absence of cost epistasis. This suggests that for cost parameters where this is the case, assuming no epistasis cannot be appropriate. Although the specific parameter values the authors use don't give rise to this problem, the authors also present more general results for a model with additive costs and no epistasis, so it would be helpful to highlight this constraint somewhere.
In box 1 we have highlighted the constraint, it is that $0 \le \β_{ij}^{x} \le \β_{ab}^{x}$. This constraint applies to any form of cost.
In the case of additive costs, there are a variety of ways this constraint could be implemented. We discuss this on lines 762772 and provide two examples – one of which can lead to epistasis.
b. The authors' careful consideration of how the specification of costs affects interpretation is indeed very useful and interesting. In addition to this conceptual point, the authors say that they are aiming to highlight mechanisms – e.g. variation in transmission rate – that could plausibly give rise to patterns of MDR. In the case of equilibrium patterns, variation in transmission rate only gives rise to LD when transmission costs are additive. The plausibility of variation in transmission rate as an explanation for equilibrium patterns of MDR therefore depends on the plausibility of additive transmission costs. Therefore, if the authors want to suggest that variation in transmission rate is a plausible explanation for patterns of MDR, it is necessary to include a discussion of whether/how additive transmission costs might arise.
The magnitude of the costs and epistasis will be dictated by the biology of the system. For example, when discussing populations separated based upon capsular serotype, the structure of the capsule has a role in determining transmission/duration of carriage. If the resistance mechanism does not interact with these differences in capsule structure, then carriage of the resistance allele will come with additive costs. However, it is also possible that the differences between capsule serotype may interact with the resistance mechanism. For example, more transmissible capsular serotypes may pay lower or higher costs of resistance. If more transmissible serotypes pay higher costs, this is multiplicative costs (or at least how we have been using it); if more transmissible capsular serotypes pay lower costs, this would require a slightly different form of the costs – but this would also yield epistasis. Similar logic applies to duration of carriage. We have added a section in the Methods (lines 754761) to address this.
More generally, without specific evidence of an interaction between differences between capsular serotypes and the resistance mechanism, then a priori it is reasonable to assume there is no interaction (additive costs).
4. Add "per carriage episode" to the summary of the biological interpretation of the duration of carriage effect in Lehtinen et al., 2017/2019 (lines 244245), to make it clear that they were not suggesting that (host) populations with longer durations of pathogen carriage have greater antibiotic exposure. (This point was missed in the original review).
Whoops, yes we have added that. Thanks for pointing it out.
5. One final observation, which is not an essential revision and the authors should only address it if they think it will improve the paper. Points 2 and 3 are both related to the result that a model with a multiplicative transmission cost predicts, at equilibrium, LD between duration of carriage and resistance, but not transmission rate and resistance. Would the authors explain this in terms of variation in the density of susceptibles giving rise to the LD with duration of carriage, but this effect being offset by the epistatic interaction between transmission rate and cost of resistance leading to no LD between transmission rate and resistance? This might be interesting to explicitly discuss in the manuscript.
If we understand this scenario correctly, the reviewer is suggesting a model in which the selection coefficients are of the form :
$s_{d}^{x} = c_{\β_d}^{x}\β^{x}S^{x} + \tau_{d}^{x}$ and there is variation in duration of carriage (but no costs of resistance to duration of carriage).
In this model, there is no way to isolate the susceptible density so that it is the only thing in the selection coefficient varying across populations without also setting $\β^{x} = \β$. This is because the parameter controlling the `costs’ in this case is not just $c_{\β_d}^{x}$, but is actually the composite parameter $c_{\β_d}^{x}\β^{x}$. This is related to our point about the interpretation of multiplicative costs: they assume a dependence upon other lifehistory characteristics.
At equilibrium, as the reviewer pointed out, there is cancellation between the $\β^{x}$ term in the costs and the $\β^{x}$ term in the denominator of the susceptible density. But this is really just a product of the mathematical assumptions; and it would be incorrect to refer to c_{\β_d} as the `costs’ because the cost `parameter’ actually depends upon $\β^{x}$.
Put another way, in the most general setting we could have simply stated we have transmission rates \β_{ij}^{x} without specifying any relationship whatsoever between `costs’ and ‘transmission’; in this case it is easy to see that only if we make specific mathematical assumptions will cancellation occur.
For example, if we were interested in a scenario in which populations with higher transmission rates pay disproportionately higher costs of resistance (which is the same biologic assumption as multiplicative costs), we would be equally justified in choosing the function
\β_{Ab}^{x} = \β^{x}/(1 + c_{\β_A} \β^{x}) (*)
as we would be in choosing multiplicative costs
\β_{Ab}^{x} = (1 – c_{\β_A})\β^{x} (#)
But using the function (*), at equilibrium \β^{x} won’t disappear from the invasion condition as it will with the function (#).
We have now highlighted in Box 1 that costs need not necessarily be ‘additive’ or ‘multiplicative’; they could just as easily take a more complex form as in our example (*) above. We simply want to emphasize that the choice of costs can produce epistasis, which is an important consideration for modelling multilocus dynamics.
https://doi.org/10.7554/eLife.65645.sa2Article and author information
Author details
Funding
Natural Sciences and Engineering Research Council of Canada (Postdoctoral fellowship)
 David V McLeod
Agence Nationale de la Recherche (ANR17CE350012)
 Sylvain Gandon
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
We thank Sonja Lehtinen, Brian J Arnold and Stephen M Kissler for helpful comments. This work was supported by a NSERCCRSNG postdoctoral fellowship to DVM. SG acknowledges support from the Agence Nationale de la Recherche, grant ANR17CE350012 ‘EVOMALWILD’.
Senior and Reviewing Editor
 George H Perry, Pennsylvania State University, United States
Reviewers
 Sonja Lehtinen, ETH Zurich, Switzerland
 Stephen M Kissler, Harvard TH Chan School of Public Health, United States
Publication history
 Received: December 10, 2020
 Accepted: May 28, 2021
 Accepted Manuscript published: June 1, 2021 (version 1)
 Version of Record published: June 16, 2021 (version 2)
Copyright
© 2021, McLeod and Gandon
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics

 956
 Page views

 136
 Downloads

 5
 Citations
Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading

 Ecology
 Evolutionary Biology
Sustainable cities depend on urban forests. City treespillars of urban forests  improve our health, clean the air, store CO_{2}, and cool local temperatures. Comparatively less is known about city tree communities as ecosystems, particularly regarding spatial composition, species diversity, tree health, and the abundance of introduced species. Here, we assembled and standardized a new dataset of N=5,660,237 trees from 63 of the largest US cities with detailed information on location, health, species, and whether a species is introduced or naturally occurring (i.e., 'native'). We further designed new tools to analyze spatial clustering and the abundance of introduced species. We show that trees significantly cluster by species in 98% of cities, potentially increasing pest vulnerability (even in speciesdiverse cities). Further, introduced species significantly homogenize tree communities across cities, while naturally occurring trees (i.e., 'native' trees) comprise 0.51%87.3% (median=45.6%) of city tree populations. Introduced species are more common in drier cities, and climate also shapes tree species diversity across urban forests. Parks have greater tree species diversity than urban settings. Compared to past work which focused on canopy cover and species richness, we show the importance of analyzing spatial composition and introduced species in urban ecosystems (and we develop new tools and datasets to do so). Future work could analyze city trees and sociodemographic variables or bird, insect, and plant diversity (e.g., from citizenscience initiatives). With these tools, we may evaluate existing city trees in new, nuanced ways and design future plantings to maximize resistance to pests and climate change. We depend on city trees.

 Ecology
Are animals’ preferences determined by absolute memories for options (e.g. reward sizes) or by their remembered ranking (better/worse)? The only studies examining this question suggest humans and starlings utilise memories for both absolute and relative information. We show that bumblebees’ learned preferences are based only on memories of ordinal comparisons. A series of experiments showed that after learning to discriminate pairs of different flowers by sucrose concentration, bumblebees preferred flowers (in novel pairings) with (1) higher ranking over equal absolute reward, (2) higher ranking over higher absolute reward, and (3) identical qualitative ranking but different quantitative ranking equally. Bumblebees used absolute information in order to rank different flowers. However, additional experiments revealed that, even when ranking information was absent (i.e. bees learned one flower at a time), memories for absolute information were lost or could no longer be retrieved after at most 1 hr. Our results illuminate a divergent mechanism for bees (compared to starlings and humans) of learned preferences that may have arisen from different adaptations to their natural environment.