Closing the life cycle
To understand which parasite genotype benefits from the within-host interaction, it is necessary to define parasite fitness. An intuitive approach is to count the number of transmission forms, when they exist (e.g. in the case of Plasmodium, one can sum up the total density of gametocytes during the infection), or the parasite load relevant to transmission in systems lacking such specialised transmission forms.
In most studies, however, it is assumed that the transmission success from the studied host is equivalent to the parasite's between-host fitness calculated over the whole life cycle. This assumption is much more likely to be inadequate for parasites that are not directly transmitted; the link may be confounded by trade-offs between traits expressed at the same part of the parasite life cycle or between traits expressed at different parts of the life cycle. As an example, consider a parasite that produces a free-living stage. An appropriate measure of transmission success from the host would be the number of spores released by infected hosts over the duration of an infection. However, if there is a very strong trade-off such that the number of spores produced is inversely proportional to their ability to infect a new host or to their survival in the environment, then the number of spores produced is less likely to be an accurate predictor of which parasite wins the within-host competition. Another example is vector-borne parasites, for example, Plasmodium. As mentioned above, the total gametocyte load in the vertebrate host is a straightforward estimate of the transmission success to vectors. However, if the ability to exploit the vertebrate host (and produce gametocytes) is traded-off against the ability to exploit the mosquito vector, assessing which strain benefits from coinfections becomes less straightforward. Such trade-offs will not necessarily exist across all life-cycle stages in all systems. It is, however, important to keep in mind the entire life cycle of the parasite before making definitive claims concerning the evolution of its virulence even during a part of its life cycle.
To accurately predict parasite evolution, the within- and the between-host levels need to be analysed jointly. The study of multiple infections is deeply rooted in ecology and evolution. As perceived already by Eshel (1977), a multiple infection framework is very similar to a meta-population framework with dispersal between patches and interactions between unrelated species within a patch. Notably, these models were published before the first epidemiological models of virulence evolution with multiple infections (e.g. Levin & Pimentel 1981). An advantage of these subdivided populations models is that they can easily allow for high diversity of infections (Gandon & Michalakis 1999), but they only allow for limited epidemiological feedbacks. There are currently two theoretical approaches to studying the evolutionary consequences of multiple infections: the superinfection framework and the coinfection framework.
The superinfection framework assumes that upon infection of an already infected host, the most competitive strain (at the within-host level) ousts the other instantaneously (Fig. 3a). That stable (or long term) coexistence within the host never occurs can be seen as a strong assumption. For instance, even between bacteriophages lambda, for which superinfection was supposed to the rule, recent work has shown that there can be multiple integration sites, which would allow coexistence between two bacteriophages lambda (Fogg et al. 2010). In the simple setting of these models, a virulent strain has a benefit at the within-host level (because it is more adapted to take over new hosts or to resist superinfection). However, virulence incurs a cost at the between-host level: in single infections, virulent strains do not maximise the number of secondary infections because they kill their hosts too quickly. Overall, assuming that superinfections occur and that more virulent strains are more competitive leads to the prediction that intermediate levels of virulence are favoured. This optimal level of virulence integrates the probability for a host to be superinfected: if superinfections never occur, the optimal virulence is zero and if superinfections always occur, the optimal virulence is infinite.
The great advantage of the superinfection model lies in its simplicity: it incorporates multiple infections into a model without adding any host category (one only needs to add a transition from one type of infected host to the other). In spite of their simplicity, superinfection models already capture some of the features of more complicated models (e.g. the selection for more virulent strains if virulence confers a within-host advantage). Finally, these models are analytically tractable, which means that their behaviour is extremely well understood. They can thus help us understand the results of more complex models.
Importantly, the effect of superinfection (and of multiple infections in general) on virulence evolution is independent of the existence of a trade-off linking virulence and transmission. Indeed, even if virulence is completely independent of transmission or recovery, superinfections can lead to the selection of virulent strains if more virulent strains are more competitive at the within-host level (Alizon & Michalakis 2011). Of course, it is also possible to combine the trade-off model and superinfections (Nowak & May 1994; Gandon et al. 2002; Alizon & Michalakis 2011).
As most of the time parasite strains co-exist for at least some time within the host, an alternative approach to the superinfection framework is the ‘coinfection’ framework (Fig. 3b,c), in which parasite coexistence within a host is assumed to last until the host dies or recovers (van Baalen & Sabelis 1995a; May & Nowak 1995). In most coinfection models, competitive exclusion cannot occur in coinfected hosts, which is one of the reasons why these models have been argued to be unrealistic. However, this competitive exclusion could be modelled, for example, by allowing for partial recovery (i.e. clearance of only one of the strains) and having this clearance rate be a function of strain within-host competitive ability. Alizon & van Baalen (2008) also used a nested model, which combines within-host and epidemiological dynamics, to account for this strain replacement in coinfected hosts. They showed that this can alter the evolutionary outcome of the model by generating an ‘evolutionary branching’ in the parasite population, that is, the evolution of a dimorphic parasite population with one morph being adapted to infecting susceptible hosts and another morph adapted to infecting already infected hosts. Note that similar results have been obtained in superinfection models (Gandon et al. 2002). Coombs et al. (2007) also built such a nested model for a case where the connecting parasites are always cotransmitted. They find the classical result that the strain that has the optimal within-host strategy can be outcompeted at the between-host level, and they also show that allowing for mutation from one strain to the other strongly affects the behaviour of the model since all hosts are infected by the two strains.
Coinfection models are more difficult to analyse than superinfection models (they always require at least an additional host class), but they allow incorporating more biology. Moreover, accounting for coinfected hosts incites us to specifying the transmission rate from each of the coinfecting strains and the overall virulence. One of the results of these models is that the type of overall virulence can indirectly affect the optimal level of virulence (Alizon 2008). This is where the knowledge from experimental studies (Table 2) proves to be crucial. Also, explicitly tracking the prevalence of coinfected hosts is useful: the outcome of within-host competition is obviously unimportant if coinfected hosts are extremely rare, but it becomes crucial if coinfected hosts are common. We already mentioned the fact that the increase in virulence due to within-host competition should lead to a decrease in the prevalence of coinfections and hence to a decrease in virulence (van Baalen & Sabelis 1995b). As a result of this feedback, the strains should evolve towards an intermediate level of virulence. Depending on the sensitivity of hosts to coinfection, that is, their probability to be infected when they are already infected (Alizon & van Baalen 2008), strains will adapt more to the within-host level (which for this type of within-host interactions means more virulent) or to the between-host level (less virulent). More recently, Alizon & Lion (2011) showed using the model of bacteria producing siderophores that epidemiological feedbacks can alter the direction of virulence evolution. Indeed, classical models that focus on the within-host level predict that increased levels of coinfection should select for less virulent strains (i.e. ‘cheating’ strains that do not produce siderophores). When epidemiological feedbacks are accounted for, the relationship between prevalence of coinfections and virulence can become non-monotonic such that, at first, increasing the average number of strains per host selects for more virulent strains but beyond a given number of strains further increasing the number of strains actually favours less virulent strains.
Importantly, there are two main families of coinfection models that make different assumptions regarding the differences between coinfecting strains. For clarity, we refer to these as coinfections by different strains of the same species (van Baalen & Sabelis 1995a) and coinfections between strains from different parasite species (Choisy & de Roode 2010). The latter model makes no assumption about the resemblance between the coinfecting strains. This is why it assumes two resident strains [one for each species, (Fig. 3c)]. The downside of the ‘two species’ coinfection model is that it requires numerical approaches because equilibrium densities of the resident state cannot be computed analytically. Conversely, the ‘one species’ model assumes that the coinfecting strains are very similar. In a way, this model can be seen as a version of the two-species model in which the two species have so few differences that all the hosts singly infected by one of the two resident strains can be grouped in the same class. This ‘one species’ evolutionary epidemiology model can seem somehow counter-intuitive because it requires to account for hosts infected twice by the resident strain. Indeed, as first noticed by van Baalen & Sabelis (1995a), without this additional category, the mutant strain always has an advantage because of its rarity, which means it will always grow (see also Alizon 2008; Lipsitch et al. 2009). If one assumes that hosts coinfected by the same strain are identical to singly infected hosts (except that they cannot be infected), which makes sense for micro-parasites, then this assumption becomes less important.
Overall, the ‘two-species’ model is never wrong but it can be unnecessarily complicated. More analyses are needed to determine the usefulness of this model, but the most promising ideas seem to apply it to specific diseases, especially if the two species are known to interact within the host in a specific way (e.g. if one species facilitates the infection by another species).
As pointed out above, the particular mechanisms by which parasites interact within hosts can have important effects on the evolution of parasite virulence. What is currently less clear is how virulence evolves when multiple interactions occur at the same time, as is likely to be the case in natural systems. Thus, experiments, which utilise natural systems and make no assumptions about which processes occur, are necessary to determine how multiple infections drive the evolution of virulence. Furthermore, as we have explained, single generation experiments are useful for determining the processes by which parasites interact, but do not lead to firm conclusions about the direction or magnitude of the evolution or virulence. Therefore, we suggest that experimental evolution studies are the best way to determine how multiple infections drive the evolution of virulence.
The first step in such an approach would be to find a suitable host–parasite system. The second step would be to determine whether parasite isolates with variable levels of virulence are available. An advantage of using an experimental evolution approach is that distinguishing between strains on the basis of molecular, immunological or morphological markers is not necessary (one can compare the evolutionary outcomes obtained with or without coinfections). Since different parasite species are often easier to distinguish than different strains of the same species, studying multiple infections by different species is likely to be easier than those by multiple strains of the same species. Furthermore, accurately controlling for the prevalence of confections by parasites from the same species seems unfeasible (one can vary host density to vary the prevalence of coinfections but this is not a precise control and it might introduce complex epidemiological feedbacks). The third step would be to determine how the variable strains or species interact within the host: for example, do the strains compete over host resources, or is there public good production? This third step is not actually mandatory since the result of the experimental evolution is interesting in itself.
Although this outline for setting up an experimental evolution approach is simplified, it is not unrealistic. In fact, Garbutt et al. (2011) studied the evolution of the bacterial insect parasite Bacillus thuringiensis (Bt) in SPE. They showed that when a pathogenic Bt strain was passaged alone it evolved towards higher levels of virulence than when it was passaged in coinfection with a non-pathogenic Bt strain. However, it is difficult to compare the outcome of this experiment to existing theoretical models because the within-host interactions between these two bacteria involve both PG production (virulence factors that kill the host) and interference (production of bacteriocins that kill unrelated strains), a situation not yet theoretically analysed. Furthermore, the fact that the bacteria are manually passaged from one host to the next also affects virulence evolution in a way that is not included in classical models (Ebert 1998). Completing the life cycle does not necessarily address the concern about the importance of epidemiological feedbacks, and in particular the fact that not all hosts are coinfected (or even infected). Again, a solution could come from an experimental approach where the infections would occur ‘naturally’. The study discussed above used SPE but letting the parasite evolve in a host population would allow for epidemiological dynamics. These dynamics might differ from those observed in vivo, especially if new susceptible hosts are added regularly to the medium. However, any dynamics are better than no dynamics because it is important that hosts do not all have the same state. Furthermore, it is also possible to build adequate epidemiological models that capture this setting.
Species or strains?
Does it matter whether parasites competing within the same host belong to different species or are different genotypes of the same species? In both cases, we are interested in the evolution of virulence within a species, so the question addresses the issue of whether selection on virulence through coinfection is exerted by other genotypes of the same species, or by genotypes of another species.
Intuitively, different parasite species are more likely to differ than different genotypes of the same species in many respects: the types of resources they extract from the host, the types of host tissues they exploit, the type of immune defences they elicit and the extent to which these immune defenses cross-react, the transmission mode, the way virulence is linked to between-hosts transmission. It is thus probable that the within-host mechanisms involved differ between the two cases. However, precisely what matters most is knowing the nature and magnitude of the effects of the different mechanisms in operation.
For clarity, in the two-‘species’ model, the two competing genotypes might very well belong to the same species. In fact, if the within-host interactions are the same, the evolutionary stable strategy (ESS) for the parasite in a one-‘species’ model should be identical to the co-ESS reached by the two ‘species’ in a two-‘species’ model. One of the reasons for this lack of difference in the predictions (and also this interchangeability of the terms strains or species in the model) is that models of virulence evolution typically assume that parasites are asexual. This state probably partly reflects the fact that we are clueless about the genetic basis of virulence (sensu decrease of host fitness), one of the reasons being that it depends on interactions between the host genotype, the genotype of both parasites and the environment (i.e. G × G × G × E interactions, Seppälä et al. 2009, 2012), and therefore rather than building very complicated models assuming a polygenic genetic basis of virulence ungrounded on empirical results it is better to build simple models. Moreover, even though processes, such as recombination, for example, in the case of Plasmodium (Conway et al. 1999) or HIV (Fang et al. 2004), or reassortment, for example, in the case of influenza viruses (McHardy & Adams 2009), are clearly involved in the generation of parasite diversity, there is no clear theory as to how these processes would matter to virulence, for example, in all the examples listed in Table 1. The three eukaryotes listed in Table 1 have one host where they are asexual and another host where they are sexual, and all the studies reported in the table focussed on the ‘asexual’ host.
In the end, the choice of the framework depends on our ability to distinguish different genotypes in the parasite population and/or to the knowledge of within-host interactions. If the two parasites are indistinguishable, we can use a ‘two species’ coinfection model but since there would be no support for differences between the ‘species’, all singly infected hosts would behave in the same way and we would de facto end up with a one-species model. Of course, the research goal matters and even if there are no known differences among the parasites, the two-species model can be used to explore potential consequences of differences between parasite species.
When performing experiments, the key point is the ability to sort different genotypes, because it allows comparing treatments with or without coinfections. The knowledge of within-host interactions in coinfections is also important but without the ability to sort out genotypes, it becomes difficult to have a ‘reference’ treatment. There are some possibilities, such as varying the input of susceptible hosts to vary the average number of genotypes per host, but this also generates complicated epidemiological feedbacks (Day & Gandon 2007).