Abstract
The bacterial pathogen Streptococcus pneumoniae is a major public health concern, being responsible for more than 1.5 million deaths annually through pneumonia, meningitis and septicemia. In spite of vaccination efforts, pneumococcal carriage and disease remain high, since available vaccines target only a subset of serotypes and vaccination is often accompanied by a rise in non-vaccine serotypes. Epidemiological studies suggest that such a change in serotype frequencies is often coupled with an increase of antibiotic resistance among non-vaccine serotypes. Building on previous multi-locus models for bacterial pathogen population structure, we have developed a theoretical framework incorporating variation in serotype and antibiotic resistance to examine how their associations may be affected by vaccination. Using this framework, we find that vaccination can result in rapid increase in frequency of pre-existing resistant variants of non-vaccine serotypes due to the removal of competition from vaccine serotypes.
Introduction
The bacterial pathogen, Streptococcus pneumoniae (or the pneumococcus) is estimated to be responsible for a third of all pneumonia cases and annually causing tens of millions of severe infections worldwide [1]. Two major tools are available for reducing pneumococcal disease: antibiotic treatment, and vaccination [2]. Antibiotic drugs remain an efficient way of clearing pneumococcal infections, but their efficiency is often impaired by the emergence and spread of antibiotic resistant pneumococci [3]. Vaccination can prevent new pneumococcal infections by increasing host immunity against pneumococcal serotypes, and eventually substantially reduce the incidence of infection by creating herd immunity [4, 5]. However, available vaccines, such as the pneumococcal conjugate vaccine (PCV), target only a subset of circulating pneumococcal serotypes, hence exerting selective pressure which can shift serotype frequencies – a phenomenon termed Vaccine induced Serotype Replacement (VISR) [6-8]. Changes in genetic composition of pneumococci have also been observed after vaccination [9, 10], and it has been proposed that vaccination may induce a shift in metabolic profiles of non-vaccine strains (known as Vaccine Induced Metabolic Shift, or VIMS), as a consequence of resource competition amongst bacteria sharing the same metabolic alleles [11, 12].
The deployment of pneumococcal vaccines has also led to significant changes in antibiotic resistance frequencies. It is unsurprising that vaccines targeting pneumococcal serotypes that have high resistance frequencies would lead to the reduction of resistance at a population level [13-17]. However it is not easy to account for a post-vaccination increase in resistance frequencies of subsets of non-vaccine type (nVT) pneumococci, as has been repeatedly observed in a range of locations [18-26]. Interestingly, resistance frequencies have not changed equally between all serotypes; and, within serotypes, resistance to different antibiotics has not changed uniformly [18-26]. Figure 1A shows changes in antibiotic resistance within the 19A serotype, an nVT of PCV7, following the introduction of this vaccine in 2000 within pneumococcal isolates of children ≤ 5 years collected in Massachusetts [10]; we observe an increase in the minimum inhibitory concentration (MIC) to erythromycin between 2001 and 2007, whereas resistance to penicillin and ceftriaxone remains unchanged (Figure 1A). It should be noted that the increase in erythromycin resistance occurred in spite of a decline of antibiotic prescription in the population over the study period [27].
A Bayesian Analysis of Population Structure (BAPS) revealed significant admixture of resistance-associated alleles from various different serotypes (Figure 1B; see Methods for full description). This analysis showed that only 87% of antibiotic resistance-associated allele combinations within serotype 19A were found to have originated from the 19A serotype itself, with the rest much more likely to be found in other serotypes. This exemplifies that 19A strains, notorious for increasing in resistance post-vaccination [18, 28-30], may experience changing selective pressures leading to changes in their distribution post-vaccination. The analysis was repeated for another publicly available UK pneumococcal data set [31] wherein the 19A serotype had the second highest value of antibiotic resistance alleles most likely originating from other serotypes (Supplementary Information Figure S1).
Here we explore the interactive evolution of vaccination and antibiotic resistance using a multi-locus model of serotype and antibiotic resistance with full flexibility in associations between alleles, reflecting the high levels of admixture shown above. In contrast with previous multi-locus models of pneumococcal evolution [11, 12] which assume interference occurs between organisms carrying similar metabolic and virulence alleles, we assume that antibiotic resistant strains are less likely to co-infect individuals infected with a susceptible strain due to ecological competition. We find that a post-vaccination surge in antibiotic resistance frequencies can occur in nVTs under these circumstances and, furthermore, may be hastened by asymmetries in rates of acquisition of resistance to different antibiotics.
Results
Model structure
We investigate the impact of vaccination within a system containing two streptococcal serotypes, a and b, of which the first is included in the vaccine (VT) and the second is not (nVT). We assume that immunity is serotype-specific, but may be incomplete, with its efficacy represented in our model by the parameter 0≤ γ ≤1, where γ=1 implies that immunity is complete and γ=0 corresponds to no serotype-specific immunity. We denote a bacterial strain of serotype a or b as having a resistance profile j, where j takes values in {00,01,10}, corresponding to sensitivity to both antibiotics, resistance to antibiotic X, and resistance to antibiotic Y.
The intrinsic transmissibility of a strain can be represented by its basic reproductive number [32], R0, which is a product of the duration of infection (D) and degree of infectivity (β), where the latter is effectively a combination of parameters defining the likelihood of acquisition by a susceptible individual of a particular strain from an infected individual. We assume that the cost of resistance would typically translate into lower infectivity of resistant strains compared to sensitive strains (β00 > β 01); however, the duration of carriage may be longer for resistant strains due to antibiotic usage (D01>D00). Thus, in the absence of antibiotic usage, the basic reproductive number of resistant strains, R010, will typically be lower than the than the basic reproductive number of sensitive strains, R000, but this can be reversed with antibiotic usage. We further assume that intrinsic fitness differences (such as in growth rates) between resistant and susceptible strains may allow an individual carrying a susceptible strain of pneumococci to suppress co-infection by a resistant strain to a degree 0≤ ψ ≤1. Note that this is a form of ecological competition between bacterial strains and is not mediated by immunity: thus individuals carrying strain a00 may not be available for co-infection by either a01 or b01(for example if ψ = 1) but will be fully susceptible to further infection by b00.
Vaccination can increase the frequency of antibiotic resistance among nVTs
We start by considering a model in which the two serotypes, a and b, are either susceptible or resistant to a single antibiotic (Y) and there are no serotype-specific differences in R0. The equilibrium frequencies of the four strains (a00, a01, b00, b01) before vaccination are determined by the degree of serotype-specific immunity (γ) and inhibition of co-infection by resistant strains (ψ) within the system, and the basic reproductive numbers of sensitive and resistant strains (Rs0 and Rr respectively). High serotype-specific immunity leads to the competitive exclusion within a given serotype, typically of the strain with the lower R0 [33]; however, the inhibition of co-infection by susceptible strains (ψ > 0) places a further cost on resistant strains such that they may be excluded even if they have a slightly higher R0. Under these circumstances, the removal of serotype a (the VT) through vaccination may cause a reversal of the outcome of competition between b00 and b01, with the resistant nVT completely replacing the susceptible nVT, as shown in Fig 2A. This is due to the removal of ecological competition between a00 and b01, and can be observed under circumstances where R01 is somewhat in excess of R00, provided ψ is above the following threshold (see Supplementary Information S2 for derivation):
When serotype-specific immunity is complete (γ = 1), the region where a surge in frequency of the resistant nVT occurs is limited to the area between Rr0 > Rs0 and the curve described by eqn (1). When R010 <R000, there is no change in the outcome of within-serotype competition as the susceptible NVT will continue to dominate after the VT is removed (Figure 2C, left of the dashed line); when R010 is sufficiently in excess of R000, b01 will have displaced b00 prior to vaccination and no change will be seen (Figure 2C, right of the solid black curve).
At lower levels of serotype-specific immunity (γ < 1), susceptible and resistant strains may coexist within the same serotype within certain boundaries of difference in R0: under these circumstances, vaccination can cause a surge in the frequency of the resistant nVT even when R010 <R000 (Figure 2 D&E) since this does not invariably lead to the total exclusion of b01. In this case, the resistant strain remains the rarer strain post-vaccination but may substantially increase in frequency, as shown in Figure 2B. In the region R010>R000, eqn (1) still determines whether the resistant nVT will increase from being the rarer strain before vaccination to being the more common strain following vaccination (Supplementary Information 2). To the right of the curve, b01 is already the more frequent strain before vaccination but may increase in prevalence after vaccination due to the cessation of competition from a00 (Figure 2 D&E).
Increasing the R0 of serotype a leads to an expansion in the parameter range within which there is a post-vaccination surge in b01 (Figure. 2 F-H). This is because the prevalence of a00 increases, causing b01 to be suppressed further; thus b01 experiences a greater increase in frequency when a00 is removed by vaccination Importantly, the general principles illustrated above remain unaltered when we introduce a second antibiotic, X, with pre-vaccination equilibria falling into 3 categories (i) coexistence of all strains (ii) competitive exclusion of both antibiotic resistant strains (b10 and b01) (iii) competitive exclusion of the resistant strain with the lower R0 (say b10). Following the removal of serotype a through vaccination, resistant strains that are already present will increase in frequency and there may be an emergence of strains that were excluded in the pre-vaccine era (Figure S2).
Effect of asymmetries in rates of acquisition of resistance to different antibiotics
The model can be extended to explicitly incorporate rates of acquisition of resistance to two antibiotics, X and Y, by introducing the parameter ωj to describe the probability of a sensitive strain acquiring a resistance profile j. We find that this has a significant impact on the outcome of vaccination where the less transmissible strain (in this case, b10, which is resistant to X) is associated with a higher rate of resistance acquisition (ω01 <ω10,). Under these circumstances, b10 may stabilise at higher pre-vaccination frequencies to b01, despite a significant transmissibility disadvantage (in Fig 3A, R010=0.91R001). The removal of competition from a00 due to vaccination effectively unmasks the transmission advantage of b01, thereby driving a rapid increase in its frequency. A stochastic implementation of this model indicates that a very significant rise in frequency of b01 can occur within a decade or two after the removal of the VT (Figure 3B) under realistic parameter combinations, and that it has a strong likelihood of eventually displacing b00 as the dominant strain within this serotype. Moreover, the time point when b01 first becomes more common than b10 is likely to be even within a decade from vaccination, but it is possible for it to occur after more than 50 years (Figure 1C). Therefore, substantial variance in surges of antibiotic resistant nVT strains post-vaccination is expected even between populations experiencing similar conditions.
As might be expected, increasing ω10 lowers the pre-vaccine frequency of b01, leading to a higher post-vaccination surge (Figure 3D). The increase in b01 is also more pronounced at higher values of ecological interference from a00 (ψ) (Figure 3 D & E), in line with our previous results. The increase in resistance among nVTs also depends on the strength of serotype-specific immunity (γ), as this determines the extent to which b01 can realise its transmission advantage (Figure 3E).
Discussion
Understanding the population dynamics of Streptococcus Pneumoniae is an important endeavour from a public health perspective, and the post-vaccination surge in antibiotic resistance frequencies observed in some nVTS is of special concern.
Existing models of antibiotic resistance typically aim to define the conditions minimizing resistance emergence or spread under different antibiotic regimes [34-37]. Fewer efforts have been made to study the effects of pneumococcal vaccination on the evolution of antibiotic resistance [38-40] and only in one of these (Lehtinen et al [40]), as far as we are aware, has a mechanism been proposed by which vaccination may induce an increase in antibiotic resistance.
A crucial difference between our model and that of Lehtinen et al, is that we include serotype-specific immunity following natural infection, although by no means does this have to be completely sterilising. Indeed, stable coexistence of resistant and susceptible strains in the pre-vaccine era is more likely to occur, in our model, under incomplete immunity. In the absence of serotype-specific immunity, coexistence becomes difficult to obtain: Lehtinen et al. show, however, that heterogeneity in duration of carriage can maintain coexistence. Within their framework, the removal of vaccine strains permits longer duration of carriage, thereby leading to an increase in antibiotic resistance provided the associated genes are in epistasis with genes influencing carriage duration; Lehtinen et al. provide evidence implying such an epistatic interaction by analysis of carriage duration and antibiotic resistance from observed data.
Our model also relies on the removal of ecological interference from vaccine strains, but here this alters the outcome of competition between resistant and sensitive strains within an nVT, potentially leading to a surge in frequency of the resistant nVT. This provides a possible explanation to the surge in antibiotic resistance in nVTs, observed independently in various populations [18-26]. We account for scenarios facilitating a surge in frequency of all or only a subset of the resistant types, under coexistence or competitive exclusion, and for varying transmissibility of the different strains. Furthermore, we introduce the notion that a resistant nVT can increase even more in frequency following vaccination if it were masked by a low rate of acquisition of resistance to other antibiotics before vaccination.
We believe that further genetic and phenotypic data of pneumococci, pre- and post-vaccination, would help distinguish between these hypotheses, and eventually lead to the design of interventions that will prevent post-vaccination increases in antibiotic resistance.
Methods
Model
In our model, each strain genotype is defined by the tuple (i, j), where i determines serotype and j the antibiotic resistance allele, respectively. For the simple bi-allelic, two-locus case, let i ∈ (a, b), j ∈ (00,01). We denote by yij the proportion of individuals currently infected by strain ij; zi is the proportion of the population previously exposed to serotype i; Zi is the proportion of the population previously or currently exposed to serotype i; Yij and Vij will refer to primary and secondary infections with strain ij, respectively. For example, the proportion of individuals infected by susceptible bacteria of serotype a is ya00; the proportion individuals previously exposed to serotype a is given by za.
Let za be the proportion of individuals who have been infected with antigenic type a, and ya01 contain all individuals currently infected with a01. Let us first assume (i) that hosts infected by a bacterial strain i, j cannot be re-infected by a strain with the same serotype, i (ii) a host infected by bacteria susceptible to antibiotics can only be co-infected by susceptible strains of bacteria.
The equations of the epidemiological model are given by (full derivations are presented in Supplementary file S1):
Where σij is the rate of clearance, equivalent to the inverse of infection length Dij; μ is the host removal rate; λij is the force of infection, determined by yij βij, where βijis the transmission rate of strain ij. An analogous set of equations is given for bacteria of serotype b. The basic reproduction number for strain ij is determined by . Antibiotic resistance acquisition is also possible in our framework and will be denoted by the parameter p, determining the fraction patients acquiring antibiotic resistance instead of being cleared. When more than one locus determining antibiotic resistance is modelled, we will denote by ωl the probability of acquiring resistance to a certain profile l by the susceptible strain. When not varied, parameter values are set to and , in accordance with paediatric pneumococcal colonization [38]; p = 0.05 for resistance acquisition scenarios [41].
To relax the two assumptions introduced above, we introduce two parameters:
We represent serotype-specific immunity by 0 ≤ γ ≤ 1, where γ = 1 is equivalent to the assumption postulated above with complete specific immunity, and γ = 0 corresponds to no serotype-specific immunity.
Similarly, we introduce the parameter 0 ≤ Ψ ≤ 1 to represent the probability that an individual carrying a susceptible strain of pneumococci will suppress co-infection by a resistant strain, due to the fitness cost of antibiotic resistance.
Adding these parameters yields the new set of equations:
We can extend this model to any number of antigenic alleles, and any number of bi-allelic resistance loci (see Supplementary file S1). Vaccination is added to the model by reducing the R0 of vaccine strains by 90%.
Stochastic implementation
We developed a semi individual-based implementation of our equations, based on the Gillespie stochastic simulation algorithm (SSA) [42]. Variables representing infected hosts (yij) are explicitly modelled under the SSA framework, whereas previously infected individuals are approximated via a deterministic approach: at each newly drawn time point tn+1, the individuals previously infected with strain i are given by
Where Fi marks the differential equation defined for the deterministic dynamics of patients previously infected with and represent all values of currently and previously infected individuals at time tn; and represents all parameters defined in the deterministic model. Since the number of previously infected patients is much greater than those currently infected, it can be approximated with a deterministic equation (updated by random processes). All simulations performed with a population size of 100,000. Finally, we constrain the number of hosts infected by any of the different strains to be ≥ 1, to avoid the absorbing states of strain extinction.
Bayesian analysis of population structure (BAPS)
Two data sets of pneumococcal genomes, collected from the USA [10] and the UK [31], were annotated using the BIGSdb software and assigned alleles with the Genome Comparator tool (with ATCC 700669 pneumococcal strain as the reference genome) [43]. We examined 34 loci associated with antibiotic resistance (given in Table S1) and preformed an admixture analysis on predefined clusters based on serotypes in the BAPS 6 software [44, 45]. Only serotypes containing 15 or more samples were used, leaving us with N=514 and N=391 observations for the USA and UK data respectively. Parameters used in the software were set according to the higher accuracy recommendations given in the BAPS manual: max clusters – 50; iterations – 500; reference individuals – 200; iterations per reference individuals – 20. Admixture inclusion threshold was set to p-value <0.001.
Funding
This research was supported by an EMBO postdoctoral fellowship (UO), and the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) /ERC grant agreement no. 268904 – DIVERSITY (JL and SG).