Parasite infection induces size-dependent host dispersal: consequences for parasite persistence

Abstract

Host dispersal is now recognized as a key predictor of the landscape-level persistence and expansion of parasites. However, current theories treat post-infection dispersal propensities as a fixed trait, and the plastic nature of host's responses to parasite infection has long been underappreciated. Here, we present a mark–recapture experiment in a single host–parasite system (larval parasites of the freshwater mussel Margaritifera laevis and its salmonid fish host Oncorhynchus masou masou) and provide, to our knowledge, the first empirical evidence that parasite infection induces size-dependent host dispersal in the field. In response to parasite infection, large fish become more dispersive, whereas small fish tend to stay at the home patch. The observed plasticity in dispersal is interpretable from the viewpoint of host fitness: expected benefits (release from further infection) may exceed dispersal-associated costs for individuals with high dispersal ability (i.e. large fish) but are marginal for individuals with limited dispersal ability (i.e. small fish). Indeed, our growth analysis revealed that only small fish hosts incurred dispersal costs (reduced growth). Strikingly, our simulation study revealed that this plastic dispersal response of infected hosts substantially enhanced parasite persistence and occupancy in a spatially structured system. These results suggest that dispersal plasticity in host species is critical for understanding how parasites emerge, spatially spread, and persist in nature. Our findings provide a novel starting point for building a reliable, predictive model for parasite/disease management.

1. Introduction

The rising tide of infectious parasites has motivated parasite/disease ecologists to establish the factors that influence parasite persistence and expansion in nature [1–3]. Classic studies have explored basic rules for parasite persistence in a locally well-mixed host population and advance the concept of ‘critical community size’ (i.e. the threshold host density below which a parasite species cannot persist) [3]. More recently, however, researchers have begun to recognize the importance of large-scale spatial processes, in which host dispersal plays a pivotal role in mediating spatial expansion of locally infected host groups/parasite-contaminated habitats [4–6]. In general, parasites per se have a very limited dispersal capability [7]. Hence, host dispersal is a primary determinant for the landscape-level dynamics of spatially structured parasite populations [5–7].

Host dispersal is thought to be an effective behaviour that enables host individuals to escape from parasite-contaminated habitats [8–10]. However, the evolutionarily stable strategy [11] of infected host's dispersal depends on the cost-benefit balance of dispersal: if dispersal ensures the avoidance of further infection risks with little or no mortality, theory predicts that natural selection favours increased dispersal tendencies of infected hosts (and vice versa) [10]. Predicting such parasite-induced changes in host dispersal behaviour is a crucial issue of parasite/disease ecology, since host dispersal propensity drives the spatial spread of parasites in the landscape [5,6]. However, current theories build on the implicit assumption that dispersal changes in response to parasite infection are constant within a single host species (i.e. a fixed trait) [10]. Host populations are heterogeneous entities of individuals with varying phenotype (e.g. body size), and the net benefits of dispersal may depend on pre-infection individual status. For example, the inherent costs of dispersal, such as reduced energy reserves or survival [12], may outweigh its potential benefits if the host's performance is insufficient to survive consecutive dispersal processes (i.e. departure, transition and settlement [13]). Hence, individual-level variation probably exists among post-infection dispersal propensities (i.e. a plastic response). Nevertheless, neither theoretical nor empirical studies have explored the possibility of plastic post-infection dispersal (conditional on individual phenotype) to date, and its potential consequences for spatially structured parasite populations are virtually unknown.

Here, we investigate whether a host species shows plastic post-infection dispersal propensities by using a tractable host–parasite system: larval parasites of the Japanese freshwater mussel Margaritifera laevis and its salmonid host Oncorhynchus masou masou. As with many metazoan parasites, the life cycle of M. laevis can be divided into free-living and obligate parasitic stages (cf. [7]). The free-living animals of M. laevis are sedentary, stream-dwelling benthic organisms (mussels) broadcasting millions of larval parasites (glochidia): their larvae are obligate, external parasites on the gills of masu salmon O. m. masou (or subspecies O. m. ishikawae) [14]. After a 40–50-day parasitic period, glochidia transform into juvenile mussels [14], which recruit into existing mussel aggregations or invade into unoccupied, parasite-free habitats via host dispersal (figure 1; [15,16]). Thus, host dispersal is a key factor determining the landscape-level expansion of the parasite species [15,16].

Figure 1.

Schematic representation of the life cycle of the freshwater mussel Margaritifera laevis. Glochidia released from female mussels infect the gills of masu salmon Oncorhynchus masou masou in the local habitat (local infection process). After a 40–50-day parasitic period, juvenile mussels detach from host fish and invade into unoccupied, parasite-free habitats through host dispersal (landscape-level process).

This single host–parasite interaction (no intermediate host or direct host-to-host transmission) serves as an excellent model system to test how host individuals respond to parasite infection, for the following reasons. First, salmonid populations have a clear size structure with a competitive dominance hierarchy [17,18], and their body size is known to vary positively with total energy reserves (e.g. time until fatigue) and swimming ability [19]. Therefore, salmonid body size may be a simple, but powerful predictor of post-infection dispersal propensities. Second, Margaritiferidae employ a simple infection strategy by which drifting glochidia released from female mussels parasitize the gills of salmonid hosts [14,20]. The infection status of masu salmon (infected or uninfected) can be readily manipulated, allowing us to experimentally compare the dispersal propensity between infected and uninfected fish hosts in the field. Finally, these host and parasite species inhabit relatively small streams, where direct observation of host dispersal is highly feasible [21].

In this study, we tested the hypothesis that glochidial infection induces size-dependent dispersal in salmonid fish hosts. Specifically, we predicted that glochidial infection enhances the dispersal tendency of large fish hosts, while suppressing that of small fish. We directly compared dispersal kernels between uninfected and infected fish using data from 215 marked individuals, half of which were artificially infected with M. laevis glochidia. We also examined how host fitness (growth rate) varied with dispersal distance and tested the hypothesis that small fish host have higher costs of dispersal than large fish. Finally, to predict the consequences of the observed host dispersal on landscape-level persistence and expansion of the parasite, we carried out a simulation study in a one-dimensional landscape of 100 habitat patches.

2. Methods

(a). Study site and study species

We conducted this study in the Chitose river system, Hokkaido, Japan. In Hokkaido, glochidia of M. laevis (approx. 50 µm in shell length [14]) are released in the summer, from mid- to late-July [15,22], and infect the gills of masu salmon (figure 1). Female mussels release synchronously millions of glochidial parasites into the water column (approx. 4 million glochidia per female; [14,15]), causing extremely high prevalence of glochidial infection near dense mussel aggregations (approx. 100% for hundreds of fish hosts) [15,16]. The proportion of infected fish declines sharply with distance from the infection source [15]. The maximum lifespan of M. laevis is approximately 79 years [14]. Margaritifera laevis is the only species of freshwater mussel within the river system.

Adult masu salmon spawn in the autumn, and eggs hatch and develop into juvenile salmon (parr) by early summer. The population of available hosts during the brooding period of M. laevis (beginning in July) is composed mainly of parr (fish at age 0+), which are suitable hosts for Margaritiferidae [23,24].

We conducted a mark–recapture experiment in the Osatsu stream (42°50′ N 141°36′ E), a small tributary flowing into the Chitose river. This spring-fed stream serves as a suitable experimental venue because: (i) this system is characterized by little temporal variation in water temperature (range: 4–12°C) and discharge [25], and because (ii) M. laevis does not occur in this stream (confounding M. laevis infection can be avoided). Field surveys were approved by the Hokkaido prefecture, and all research was performed in line with the Animal Care and Use guidelines of Hokkaido University.

(b). Mark–recapture experiment

We selected a 1200 m stretch of the Osatsu stream with a 3.0–6.0 m wetted width, where local habitat conditions varied little along the stream stretch, and no apparent dispersal barriers existed. The stretch was divided into 60 capture subsections (each 20 m in length). The first capture session was carried out from 21 to 24 July, 2015, and the recapture session occurred ca. 50 days later (8–11 September 2015). This capture-recapture interval was intended to mimic the duration of glochidial infection under the summer water temperature of the Osatsu stream (approx. 12°C) [26].

During the first capture session, we collected host fish using consecutive three-pass electrofishing in each subsection [27]. We anaesthetized captured fish in a 2-phenoxyethanol solution and measured their fork length (millimetres) and wet mass (0.1 g). We batch-marked them with fluorescent visible implant elastomer tags (Northwest Marine Technologies, Shaw Island, Washington) applied to three adipose locations (behind the eye, behind the nose and the lower jaw). We used six tag colours, and the combination of tag colour and position allowed the identification of individual fish (6³ = 216 patterns). We allowed marked fish to recover in a container for 10–15 min. We did not mark 1+ fish hosts as they have been identified as unsuitable hosts for Margaritiferids [23,24].

Half the marked fish collected from each individual subsection were infected with M. laevis glochidia by placing them into an infection bath (5 l bucket, 4 × 10⁴ glochidia l⁻¹) for 30 min. The other half was kept in a ‘sham’ infection bath with no glochidia (control). This experimental design isolates infection-treatment effects from any environmental variation among subsections. We obtained a preliminary confirmation that this glochidial density provides no infection failure and natural levels of glochidial load (48 ± 29 glochidia per fish), based on the level commonly observed in the Chitose river [22,28]. We created the infection bath using fresh viable glochidia that were naturally released from a total of four gravid M. laevis females (collected every morning from a single population of the Chitose river). The averages and variance of fish body size (fork length) were almost identical between infected and uninfected fish groups (t-test, p = 0.18; mean for uninfected fish = 92.5 ± 11.2 mm, mean for infected fish = 90.5 ± 11.2 mm).

Marked fish were then released near the centre of the subsection where they were caught. We completed all procedures from 07.00 to 15.00 in light of the short longevity of glochidia [29]. We marked a total of 215 individuals.

At the recapture session, we recaptured marked fish with three-pass electrofishing in each of the subsections. The longitudinal position (recorded as subsection identity (ID), 1–60), fork length, and wet mass of all recaptured fish were recorded.

(c). Dispersal model coupled with observation process

We employed the Laplace (double exponential) kernel, which has been proven to provide adequate fits to dispersal data in various salmonid fishes [30,31]. The Laplace density function, f_L, has a symmetrical exponential decay to either side of the origin, with the inverse of parameter τ equal to the mean dispersal distance (i.e. smaller values of τ indicate greater dispersal tendency):

2.1

where x_j(i) is the distance class (i.e. distance to the downstream end of the whole study section) at the centre of recapture subsection j (subscript j(i) denotes jth subsection in which fish individual i was recaptured), and μ_i is the distance class at the centre of the capture subsection for fish individual i. It is important to note that the variance () nonlinearly increases with increasing mean dispersal distance (1/τ_i). Thus, the model can express outliers adequately (i.e. robust to outlier data), which is typical for dispersal data [30,32]. We related the mean dispersal distance to individual-level predictors with a log-link function:

2.2

where β₀ is an intercept and β₁–β₃ are standardized regression coefficients of infection status Infection_i (binary variable; infected = 1, uninfected = 0), initial body size Size_i (continuous variable; standardized with a mean 0 and s.d. 1) and their interaction. As initial body size was standardized, parameter β₀ indicates the average 1/τ of the population in a logarithmic scale. This formulation allows us to evaluate the effects of individual-level predictors on the form of dispersal kernels.

To incorporate sampling designs into the parameter inference of dispersal kernels, we modified the inference framework proposed by Pepino et al. [31]. The binary variable of capture history Y_i (Y_i = 1 if recaptured, otherwise 0) was modelled based on a Bernoulli distribution (see the electronic supplementary material for derivation):

2.3

where φ_j(i) is the section-specific probability of capture and s_i is the survival probability during the study period. For recaptured individuals, D_i is the probability that individual i moves from release point μ_i to subsection of recapture j. For unrecaptured individuals, D_i represents the probability of staying in the 1200 m study stretch:

2.4

where x_j(i),up and x_j(i),low are the distance classes at the upper and lower boundaries of the recapture subsection j for individual i, and Up and Low are the distance classes at the upper (1200 m) and lower ends (0 m) of the whole study section. Individual-specific survival probability s_i is an identifiable parameter, as we obtained an independent estimate of section-specific capture probability φ_j(i) using the three-pass depletion surveys with a Bayesian modification (see the electronic supplementary material) [27]. Survival probability was normally distributed in a logit scale: logit(s_i) ∼ Normal(logit(s_global), ), where s_global represents the mean survival probability.

However, it was impossible to determine φ_j(i) for unrecaptured individuals, since we have no information on subsection ID of recapture. Instead, for unrecaptured individuals, we estimated the weighted mean of φ_j(i) across subsections (φ_w,i) given the dispersal parameter τ_i:

2.5

and

2.6

In equation (2.6), the numerator indicates the probability of movement from release point μ_i to jth subsection (i.e. the probability of unrecaptured individual i present at jth subsection during the recapture session). The denominator (the probability of staying in the 1200 m study stretch) scales the numerator so that equals 1.0.

Vague priors were assigned to the parameters: normal distributions for β (mean = 0, variance = 10⁴), a beta distribution for s_global (shape = 1, scale = 1), and a truncated normal distribution for (mean = 0, variance = 10⁴, range = 0–100) The model was fitted to the data with JAGS v. 4.1.0 and the package ‘runjags’ [33] in R v. 3.3.1 [34]. Three Markov chain Monte Carlo (MCMC) chains were run with 9000 iterations (3000 burn-in), and 500 samples per chain were used to calculate posterior probabilities. Convergence was assessed by examining whether the R-hat indicator of each parameter approached 1 [35].

(d). Host growth analysis

We analysed factors that influence host growth using data from recaptured fish (n = 116). We estimated individual host growth G as:

2.7

where FL₀ and FL₅₀ denote fork length at capture and recapture, respectively.

We then constructed a linear mixed effect model with a random effect of initial capture subsection ID to investigate factors influencing host growth G. Host growth G was assumed to follow a normal distribution and was modelled as a function of distance moved (continuous), logarithm of fork length at initial capture (log(FL₀); continuous), infection (binary), and their two-way interactions. Continuous explanatory variables were standardized prior to the analysis (mean = 0, s.d. = 1). However, an analytical issue can arise when using G as a response variable: log(FL₀) will appear in both sides of the equation, causing spurious correlations [36]. To avoid this analytical issue, we put log(FL₀) in the response variable into the right side (i.e. offset term): log(FL₅₀) = Xβ + ɛ + log(FL₀), where X is a matrix of predictors, β is a vector of regression coefficients, and ɛ is the random effect of initial capture subsection ID.

Vague priors were assigned to regression coefficients (normal distributions: mean = 0, variance = 1000) and standard deviations of residuals and the random effect (uniform distributions: range, 0–100). Three MCMC chains were run with 15 000 iterations (5000 burn-in) using JAGS, and 500 samples per chain were used to calculate posterior probabilities. Convergence was assessed as indicated above.

(e). Simulation

To investigate the consequences of the observed dispersal probabilities on the landscape-level expansion and persistence of the parasite, we modified a simulation model described by Grant et al. [37]. In our simulation, we assumed a linear landscape of 100 habitat patches of equal quality and length (20 m). Simulation space boundaries were wrapped. We initially introduced five parasite-occupied patches into the landscape (5% occupancy), and a random, independent patch-extinction of parasite-occupied patch occurred with the probability E in each time step (i.e. transition from a parasite-occupied to parasite-free patch). After random extinction events of parasite-occupied patches, we allowed immediate (re)colonization of mussel aggregation (i.e. transition from a parasite-free to parasite-occupied patch) from other parasite-occupied patches through host dispersal (figure 1): in each time step t, all host fish at parasite-occupied patch i (N_i,t) were infected with glochidia (100% prevalence of local glochidial infection; [15,16]) and dispersed randomly based on the predefined Laplace dispersal kernels (see below). Every host disperser had information on its own position (i.e. distance [m] to the downstream end). During dispersal, infected host fish survived with the probability s (s = 0.87; see Results), and each surviving individual that reached a parasite-free patch j (i.e. between x_j(i),low – x_j(i),up) had the potential to create a new infection source (parasite-occupied patch) with the probability C. Thus, the realized per-fish colonization probability is the product of s and C. Host fish present at parasite-free patches (i.e. no mussel aggregation) were not considered as dispersal agents.

The number of host fish at patch i at time step t was drawn from a Poisson distribution as N_i,t ∼ Poisson(λ_t). This reflects real situations: that is, in each time step, newly emerged susceptible hosts (i.e. 0+ fish of the year) were randomly (re)distributed across the simulation space. The mean host density (fish/patch) at time step t, λ_t, was determined by the Ricker model [38–40]:

2.8

where r is the intrinsic population growth rate, b is the parameter that determines negative density dependence, and ɛ_t is the environmental stochasticity. In this formulation, the host carrying capacity K can be written as r/b. The parameter ɛ_t is governed by a normal distribution with a mean of 0 and variance of , and the variance parameter determines the degree of environmental stochasticity. Host population dynamics were assumed to be independent of parasite infection, as we did not find negative effects of glochidial infection on fish growth (table 2).

Table 2.

Results of the Bayesian regression model that explains host growth rate. (Posterior probability represents the proportion of parameter estimates (MCMC samples) assigned to be either negative or positive. Parameters with a posterior probability of greater than 0.95 are shown in italics. s.e.: standard error.)

effect	estimate	s.e.	posterior probability
			negative	positive
intercept	0.098	0.006	—	—
body size	−0.016	0.006	1.00	0.00
dispersal distance	−0.009	0.010	0.82	0.18
infection	0.003	0.009	0.36	0.64
size × distance	0.013	0.007	0.05	0.95
size × infection	0.008	0.009	0.19	0.81
infection × distance	−0.008	0.010	0.81	0.19

The dispersal parameter τ_i for the Laplace density function was described as follows:

2.9

where γ is the intercept, δ is the slope that determines the strength of size-dependence in host dispersal, and z_i is the standardized random variable of body size, z_i ∼ Normal(0, 1). Note that the parameters (γ and δ) can be substituted by empirical dispersal estimates in equation (2.2): log(1/τ_i) = [β₀ + β₁ · Infection_i] + [β₂ + β₃ · Infection_i] · Size_i. This representation allowed us to predict the consequences of observed dispersal patterns on parasite persistence (control [Infection = 0]: γ = β₀, δ = β₂; treatment [Infection = 1]: γ = β₀ + β₁, δ = β₂ + β₃).

We examined 16 dispersal parameter combinations (γ = 1.5, 2.0, 2.5, 3.0; δ = 0.0, 0.5, 1.0, 1.5) that cover the empirical estimates (see Results). We also set the following parameters: C = 0.005, E = 0.01, K = 4 or 8, r = 1.43 and σ_ɛ = 0.18. Per-fish colonization rate C was parameterized according to the reported mussel mortality during their early benthic life stage (approx. 99.5%) [41]. Specific estimates of annual per-patch extinction probability were not available, but descriptive evidence suggests that the parameter should be ≤0.01 [42]. The range of parameter K was determined based on the masu salmon density in the Osatsu stream (0–12 fish per 20 m subsection). We used fixed values of r and σ_ɛ based on a meta-analysis by Myers et al. [39], and the density-dependence parameter b was calculated as r/K. For a connection with traditional epidemiological parameters (e.g. the basic reproductive number R₀), see the electronic supplementary material.

We ran the simulation model for a maximum of 10 000 time steps. We obtained persistence time of the spatially structured parasite population (time to the entire extinction of parasite-occupied patch) and median proportion of parasite-occupied patch (hereafter, ‘occupancy’) during the persistent period (i.e. a period during which non-zero occupancy of parasite-occupied patch was observed). Each parameter combination was replicated 25 times. The initial mean host density λ₀ was set to be five fish/patch for all cases. All simulations were conducted in the C++ environment using the R package ‘Rcpp’ [43].

3. Results

(a). Mark–recapture study

Among 215 marked fish hosts (fork length: 62–121 mm), 116 individuals (infected fish, 54 individuals; uninfected fish, 62 individuals) were successfully recaptured and identified (no complete tag loss was observed). The average capture probability of our three-pass electrofishing φ during the recapture session was reasonably high (median: 0.93, 95% CI: 0.83–0.98).

Using the dispersal data as well as the capture history of the marked fish, we developed a Bayesian dispersal model coupled with observation processes (capture probability φ, survival probability s, and sampling designs). Estimated survival probability s_global during the mark–recapture period was 0.87 (95% CI: 0.62–0.99) with among-individuals variation () of 6.03 (95% CI: 0.84–17.18). The model also revealed that parasite infection interacted with host body size to modify host dispersal kernels (table 1). Glochidial infection influenced large fish to be more dispersive, but had opposite effects on small individuals (figure 2b). The probability of leaving behind the home subsection was high for large fish hosts (0.76; 80th percentile of body size); specifically, it was 4.3 times greater than that of small fish hosts (0.18; 20th percentile of body size). However, such clear size-dependence in dispersal was not observed for uninfected fish hosts (figure 2a). The interactive effect of host body size and parasite infection remained significant even after removing two ‘super dispersers’ (i.e. outliers; only two individuals dispersed ≥180 m; see the electronic supplementary material, table S1 and figure S1). For raw data, see the electronic supplementary material, figure S1.

Table 1.

Results of the Bayesian model that explains individual-level variation in dispersal kernel. (Posterior probability represents the proportion of parameter estimates (MCMC samples) assigned to be either negative or positive. Parameters with a posterior probability of greater than 0.95 are shown in italics. s.e.: standard error.)

effect	estimate	s.e.	posterior probability
			negative	positive
intercept (β0)	2.571	0.092	—	—
infection (β1)	0.151	0.135	0.12	0.88
body size (β2)	0.319	0.089	0.00	1.00
infection × size (β3)	0.829	0.185	0.00	1.00

Figure 2.

Plastic dispersal response of masu salmon Oncorhynchus masou masou to infection by larval parasites of Margaritifera laevis. Shaded areas with dotted lines denote average dispersal kernels. Solid and dashed lines indicate dispersal kernels for large (80th percentile body size) and small (20th percentile body size) fish hosts, respectively. Uninfected individuals showed little variation in dispersal (a). By contrast, infected individuals exhibited strong size-dependence in dispersal (b).

Growth analysis revealed that dispersal costs were size-specific. The main effect of body size and the interaction term with dispersal distance were detected with a probability of ≥0.95 (table 2). The growth of small fish hosts decreased with increasing dispersal distance (figure 3). By contrast, this pattern was less apparent for large fish hosts (figure 3). The lack of small ‘super’ dispersers probably reflects the size-specific costs of dispersal. Glochidial infection had little effect on fish growth.

Figure 3.

Size-specific effects of dispersal on host growth rates (log(FL₅₀/FL₀)). Solid and broken lines represent predicted values of the linear mixed effect model for large (80th percentile body size) and small (20th percentile body size) fish hosts, respectively. Bubbles indicate individual fish and their sizes are proportional to fork length during the first capture session (FL₀).

(b). Simulation

Dispersal plasticity had a strong positive impact on the landscape-level persistence and expansion of the parasite, especially when the host population size was large (see figure 4b,d; larger values of y-axis represent stronger size-dependence in dispersal). The observed plasticity in dispersal led to approximately four times longer persistence of spatially structured parasite population with greater occupancy (approx. 8000 time steps with approx. 30% occupancy; filled dot in figure 4b,d) compared with the weak plastic dispersal scenario (approx. 2000 time steps with approx. 6% occupancy; open dot in figure 4b,d). However, this contrast was not so clear when the host population size was small (figure 4a,c).

Figure 4.

Contour plots of simulated parasite persistence times (a,b) and occupancy (proportion of parasite-occupied patch; c,d) in a 100 patch linear landscape with host carrying capacities of 4 (a,c) and 8 fish patch⁻¹ (b,d). Brighter colours represent longer persistence time (a,b) or greater occupancy (c,d). Increasing values on the γ (x-axis) denote increasing average dispersal distance, whereas increasing values on the δ (y-axis) represent stronger size-dependence in dispersal (larger individuals become more dispersive whereas smaller individuals become less dispersive). Other parameter values were as follows: colonization rate, C = 0.005; extinction probability, E = 0.01; environmental stochasticity in host population dynamics, σ_ε = 0.18; and survival during dispersal, s = 0.87. Open and filled dots represent observed dispersal scenarios with weak (γ = β₀, δ = β₂; see figure 2a) and strong (γ = β₀ + β₁, δ = β₂ + β₃; see figure 2b) size-dependence, respectively. See table 1 for estimated parameters. (Online version in colour.)

4. Discussion

Host dispersal is now recognized as a key mediator of the landscape-level persistence and expansion of parasites [4–6]. However, the plastic nature of host dispersal responses to parasites has long been underappreciated, despite the fact that intrapopulation variation in phenotype (e.g. body size) is ubiquitous. Inherent difficulties exist in manipulating infection status and quantifying dispersal in natural settings, and these problems have hindered the progress of this research field. Here, using a tractable host–parasite system embedded in a one-dimensional landscape (i.e. a stream), our field experiment overcame these difficulties and provided, to our knowledge, the first quantitative evidence that parasite infection induces size-dependent host dispersal. Strikingly, our simulation suggested that the observed individual-level variation in host dispersal may greatly enhance parasite persistence and occupancy in a spatially structured system. This is an emergent phenomenon that cannot be understood without the inclusion of among-individuals differences in dispersal propensities. These findings provide an important insight into how parasites emerge, spatially spread, and persist in stochastic natural environments.

As M. laevis glochidia parasitism occurs mainly in the vicinity of adult mussel aggregations (infection rate drops approximately 0.20 with every hundred metres of distance from aggregation) [15], dispersal seems to be an effective measure to avoid further infection for the salmonid host. Concordant with our hypothesis, there was substantial variation in host dispersal: large fish hosts became more dispersive, whereas small individuals tended to stay in the home patch. This size-specific dispersal is interpretable from the viewpoint of host fitness. In brief, host dispersers cannot benefit from parasite-avoidance unless they survive the dispersal process and reach a new, parasite-free habitat. Our results are consistent with this intuitive prediction. Even under the influence of parasite virulence, the plentiful energy reserves of large salmonid fish may enable them to survive the risky transition process. By contrast, for small fish hosts, glochidia-induced changes (e.g. respiratory burden) [44] may lead to a failure of transition, given their presumed susceptibility to energetic and/or risk costs during dispersal (see [19] for body size effects on swimming performance). Further costs can be levied at the settlement stage, as small salmonid fish are competitively inferior in the dominance hierarchy [17,18]. Indeed, our growth analysis produced some support for this interpretation, as only small fish hosts incurred dispersal costs (reduced growth; see figure 3). Therefore, the net benefits of dispersal are expected to be higher for large fish hosts, but marginal for small fish.

Alternatively, it is possible that M. laevis glochidia actively manipulated host dispersal behaviour. However, M. laevis does not seem to have strong incentive for manipulating host dispersal, as the parasite possesses neither a complex life cycle (i.e. no secondary host) nor specific spawning habitats, both of which are often associated with active host manipulation [7]. Considering the evidence, we suggest that the observed dispersal changes may be a plastic response of salmonid fish hosts, and a ‘fixed dispersal response’ to parasites (e.g. all individuals disperse more; increase in x-axis values in figure 4) may not be the best option for the host species owing to the size-specific costs of dispersal.

Intriguingly, our simulation revealed that the host's plastic response to parasite infection has the potential to increase landscape-level parasite persistence and occupancy, provided that host carrying capacity K is high. The rationale behind this finding is the following: large host population sizes ensure a certain fraction of a host population consists of movers (large fish hosts) that allow parasites to colonize distant patches and spread spatially. Meanwhile, stayers (small fish hosts) effectively reinforce (or recolonize) adjacent patches (i.e. the mixture of individual dispersal kernels makes up a ‘fat-tailed dispersal kernel’). This result deserves further attention. If the observed host response is truly adaptive, then a host's adaptive behaviour, either avoiding infection risks or dispersal costs, may lead to undesired consequences: the parasite invades a larger fraction of the patches and persists in the spatially structured system. However, we do not have the valid evidence that dispersal plasticity maintains host fitness (i.e. adaptive), and further experimentation is needed to confirm this possibility. Future investigations addressing this issue would shed light on how host–parasite interactions are stably sustained in spatially structured systems.

Mark–recapture studies are recurrently criticized for the limited coverage of potential dispersal ranges (e.g. [45]). However, our statistical approach is robust against this problem, as we incorporated sampling designs into the dispersal parameter inference (see equation (2.3); individuals that had left the study section were taken into account). This has been shown to provide reliable dispersal parameters, especially when the length of the study stretch is more than 4 times greater than the average dispersal distance [46]. This probably holds true in our study, as the length of the whole study section (1200 m) was approximately 92 times longer than the average dispersal distance of masu salmon (exp(2.57) = 13.1 m; see table 1). Therefore, we are confident in our dispersal parameter inference.

Another potential issue would be whether our findings are applicable to horizontally transmissible systems (direct host-to-host transmission). Although our system has several comparative advantages owing to the lack of horizontal transmission (e.g. infection status is readily controllable throughout the experiment), there are certain differences in local transmission dynamics. Nevertheless, we expect that our findings may be equally important in those systems because the landscape-level expansion of horizontally transmissible parasites should also occur mainly through host dispersal. The integration of local horizontal transmission is beyond the scope of our study, but this issue may be a fruitful avenue for future theoretical research.

Although the importance of host dispersal has been increasingly appreciated in the field of infectious parasite research, researchers have failed to fully account for the heterogeneous nature of wild organisms. By combining empirical and simulated approaches, the present study provides a novel parameter (i.e. individual-level variation in host phenotype) for predicting the long-term persistence of parasites and the landscape-level expansion of parasite-contaminated habitats. As intrapopulation heterogeneity of phenotype can be found in almost all animals, our findings may be widely applicable to other host–parasite systems. Future generalization across systems should provide a novel and critical perspective on parasite/disease management issues.

Ethics

Field surveys were approved by the Hokkaido prefecture, and all research was performed in line with the Animal Care and Use guidelines of Hokkaido University.

Data accessibility

Data and scripts (JAGS and Rcpp) are available at Dryad (http://dx.doi.org/10.5061/dryad.14mt6) [47].

Authors' contributions

A.T. and K.O. designed and conducted the experiment. A.T. performed statistical analysis and simulation. All authors participated in conception, discussion of the results and manuscript preparation.

Competing interests

We declare we have no competing interests.

Funding

We received no funding for this study.