Decomposing parasite fitness in a two-host, two-parasite system reveals the underpinnings of parasite specialization

The ecological specialization of parasites – whether they can obtain high fitness on very few or very many different host species – is a determining feature of their ecology. In order to properly assess specialization, it is imperative to measure parasite fitness across host species; to understand its origins, fitness must be decomposed into the underlying traits. Despite the omnipresence of parasites with multiple hosts, very few studies assess and decompose their specialization in this way. To bridge this gap, we quantified the infectivity, virulence, and transmission rate of two parasites, the horizontally transmitted microsporidians Anostracospora rigaudi and Enterocytospora artemiae, in their natural hosts, the brine shrimp Artemia parthenogenetica and Artemia franciscana. Our results demonstrate that each parasite performs well on one of the two host species (A. rigaudi on A. parthenogenetica, and E. artemiae on A. franciscana), and poorly on the other. This partial specialization is driven by high infectivity and transmission rates in the preferred host, and is associated with maladaptive virulence and large costs of resistance in the other. Our study represents a rare empirical contribution to the study of parasite evolution in multi-host systems, highlighting the negative effects of under- and over-exploitation when adapting to multiple hosts.


INTRODUCTION
replicates for the doses 400, 3 200 and 6 400 spores per 174 individual, respectively). All hosts were ~4 weeks old and measured between 5 and 8 mm; A. 175 franciscana hosts were mixed males and females. 176 Statistical analyses 177 To analyze the dose-response curves, we used four-parameter log-logistic modeling in R 178 (package drc, Ritz and Strebig 2005, R Core Team 2014). In these models, the four 179 parameters determining the shape of the sigmoidal curve are: the lower limit (set to 0 in our 180 case), the upper limit, the slope around the point of inflection, and the point of inflection 181 (which here is the same as the ED50). The (binomial) response variable was the number of 182 individuals that were infected vs. uninfected. Because we did not perform the A. 183 parthenogenetica and A. franciscana experiments at the same time, we could not control for 184 environmental effects. Thus, we simply tested if the dose-response curves for A. rigaudi and 185 E. artemiae were different within each host species. To do this, we fit models that did or did 186 not include a 'microsporidian species' effect and compared the two using a likelihood ratio 187 test. If the effect was significant, we went on to compare the parameters of the two resulting 188 curves ('compParm' function in the drc package). 189 Experiment 2: Virulence and transmission 190 Experimental design and execution 191 To quantify the virulence and transmission rates of A. rigaudi and E. artemiae, we 192 experimentally infected individual Artemia with controlled spore doses. We then tracked their 193 survival, growth, reproductive output, and spore production over a two-month period. We 194 also quantified host-to-host transmission at two time points. was changed every five days. Hosts were fed 0.5 mL algal solution daily (2.6*10 9 T. chuii 221 cells/L deionized water); this feeding regime corresponds to half of the maximum ingestion 222 rate of an adult Artemia (Reeve 1963) and has been shown to reveal energetic trade-offs 223 (Rode et al. 2011). We ended our experiment after 60 days, at which point surviving 224 individuals were sacrificed and tested for infection by PCR (following Rode et al. 2013a). 225 To quantify the effects of infection on the hosts, we tracked the growth, survival, and 226 reproduction of the experimental individuals. Body length was recorded on days 30 and 60. therefore, clutches of nauplii were counted five days after sighting. During these five days 237 nauplii were in competition for resources with their mother (plus an additional male if A. 238 franciscana, see below). However, mothers were removed at each water change, which could 239 happen before the clutch had reached the five-day mark. In these cases, we placed a new tube 240 containing one (or two, if A. franciscana) adult male Artemia above the nauplii to ensure the 241 same level of food competition. 242 While A. parthenogenetica females reproduce in isolation, A. franciscana females need to be 243 fertilized before each clutch (Bowen 1962). We therefore added mature A. franciscana males 244 from parasite-free lab stocks to each tube containing an A. franciscana female. To prevent cross-contamination between the male and the female, exposed males were removed and new 246 uninfected males added every five days (five-day estimate based on infection detection time 247 as found by Rode et al. 2013a). Male Artemia mate-guard by clasping females around the 248 abdomen (Bowen 1962), and forcible removal may be harmful to both partners. To avoid 249 this, males found mate-guarding on the fifth day were given up to two extra days with the 250 female, after which they were forcibly removed. Couples were fed twice the individual food 251 allocation.

252
To estimate parasite fitness, we estimated spore production at regular points throughout the 253 experiment. To do this, we collected 1 mL of feces (containing parasite spores) from every 254 experimentally infected host at every water change. Samples were stored in 1.2 mL Qiagen 255 Collection Microtubes and refrigerated until the spore concentration could be quantified. To 256 measure the spore concentration, we homogenized and stained each sample as described 257 above, with minor differences in the centrifugation steps (16 min at 5 000 g) and the final 258 concentration (concentrated to 14.3X by removing 930 µL of the supernate). Spores were 259 counted once per sample, as described above. Because counting spores is labor-intensive, we 260 restricted our efforts to the feces samples collected on days 15, 30, 45 and 60. 261 We also investigated the host-to-host transmission success of the parasites and its relation to 262 spore production. On days 30 and 60, we allowed a subset of experimental hosts (hereafter 263 the 'donors') to infect groups of uninfected 'recipient' hosts for 24 hours. Donors were first 264 placed with either eight A. franciscana or eight A. parthenogenetica recipients; after 24 265 hours, the donor was removed and placed with a new group of recipients of the other species.

266
All recipient hosts were taken at random from uninfected lab stocks of varying ages and sizes 267 (min = 4 mm, max = 10 mm). The donor host was separated from the recipients by a 1x1 mm 268 net; recipients swam underneath them in 40 mL of brine. Infection was allowed to incubate in 269 the recipients for six days after the donor was removed; surviving recipients were then sacrificed and PCR-tested for infection (following Rode et al. 2013a). The prevalence of 271 infection in recipient individuals could then be compared to the number of spores counted in 272 the feces samples on day 30 or 60.

273
A key aspect of infection follow-up experiments is knowing which individuals were infected 274 after exposure to the parasite, and which were not. In our experiment, testing by PCR was 275 often not sufficient to determine if an individual was infected, because individuals that died 276 before day 60 often had quickly decaying corpses and thus degraded DNA. We therefore 277 considered that an individual was infected if it tested positive by PCR or produced spores or 278 transmitted the infection to a recipient host. If none of these requirements were met, we 279 considered that the individual was not infected. By applying these criteria, we could be sure 280 of the infection status for almost all individuals that died on or after day 15 (the first spore 281 collection date); for any individuals who died before day 15 and who tested negative by PCR, 282 we could not exclude the possibility that they were infected. 284 We analyzed the results of this experiment in two major parts. First, we examined the 285 virulence of infections (effect of the parasite on host survival, growth, reproduction, and 286 overall fitness). In these analyses, we excluded all individuals that did not become infected 287 after exposure to the parasite. We also excluded all individuals that died before day 15 (we 288 could not be certain of infection status before this day, see above). To make sure that we were 289 not missing important events occurring before this cutoff, we repeated all statistical models 290 for exposed vs. control individuals that died before day 15. Second, we analyzed parasite 291 transmission (spore production rate, infectiousness, and overall fitness). An overview of the 292 analyses is given in To estimate the infectiousness of a single spore, we used the results of the transmission assay. 299 We assumed that the establishment of microsporidian infections follows an independent-300 action model with birth-death processes. This model assumes that a parasite population grows 301 in the host until it reaches an infective threshold, at which point the infection is considered to 302 be established (Schmid-Hempel 2011 pp. 225-6). In our assay, we considered that an 303 infection was established when we could detect it; in other words, the infective threshold 304 corresponded to the threshold for PCR detection (estimated at ~1 000 spores inside the host's 305 body, unpublished data). In these models, the probability per spore to start an infection, p, is 306 equal to − ln ( ) / where D is the spore dose (Schmid-Hempel 2011 pp. 307 225-6). In our transmission assay, D can be approximated by the number of spores in the 308 fecal sample taken from the donor at the start of the assay (= spore count transformed to 309 spores/mL, or * 700), divided by 5*8 = 40 (fecal samples accumulated over a 5-day period but 310 we only exposed recipients for one day; the inoculum was shared amongst 8 recipients). We 311 calculated a value of p for every replicate in the transmission assay.

312
For each infection we used two measures of spore production as proxies for parasite fitness. In most of the experimental host-parasite combinations, a subset of exposed hosts did not 328 become (detectably) infected. Hereafter, we refer to these individuals as resistant, because we 329 found a posteriori differences in the proportion of such individuals across host-parasite 330 combinations, and in their life history traits compared to infected individuals and controls. As 331 above, the analyses of these two aspects excluded all individuals who died before infection 332 status could be definitively determined, i.e. those that died before day 15 of the experiment.

333
We analyzed the distribution of resistance across host-parasite combinations. reproduction. These results are summarized in Fig. 2 and the significance of tested effects is 372 listed in Table 4; we discuss the effects of infection in more detail below. Here, we report 373 only the analyses for infected vs. control individuals, which excluded all individuals that died 374 before day 15. When we compared exposed vs. control individuals that died before the cut-375 off day the results were not qualitatively different.

376
In most host-parasite combinations, survival was reduced (Fig. 2 Table 4). The time until maturity, which was best described by a lognormal distribution 399 (ΔAICc ≥ 4.7), was significantly delayed by infection with either parasite species (post-hoc  Table 4).

425
Transmission and fitness of infections 426 Second, we studied the effects of the host species on the parasite's transmission and fitness 427 (summarized in Fig. 4). These analyses were combined for all host-parasite combinations, but 428 A. parthenogenetica that were exposed to 10 000 E. artemiae spores were analyzed separately 429 unless otherwise specified.

430
The infectiousness of a single spore (the probability that it started a detectable infection, as 431 calculated using the transmission data) corresponded with our expectations based on 432 Experiment 1 (Fig. 4). Host-parasite combination had a significant effect on infectiousness 433 (χ 2 (1) = 16.7, p < 0.0001). A. rigaudi tended to be more infectious to A. parthenogenetica than to A. franciscana (post-hoc z = 2.3, p = 0.10); E. artemiae was significantly more infectious to 435 A. franciscana than to A. parthenogenetica (post-hoc z = 3.6, p < 0.01).

441
As expected, host-to-host transmission success increased with the rate of spore production in  artemiae (low dose), so we continued our analyses with these combinations.

463
For both A. franciscana exposed to A. rigaudi and A. parthenogenetica exposed to a low  advantageously coupled with high rates of spore production in its matched host A. 508 parthenogenetica, and disadvantageously coupled with low rates of spore production in its 509 mismatched host. E. artemiae is avirulent in its mismatched host, which at first glance 510 appears ideal. However, when spore production is taken into account, it becomes clear that 511 this avirulence in A. parthenogenetica is coupled with very low rates of transmission, 512 whereas the rate of spore production is high in A. franciscana.

513
Despite this, were we to consider the component traits individually, they would not all lead us 514 to conclude that the two parasites are partially specialized. The pattern of spore production in 515 the matched vs. mismatched combinations best reflects the overall degree of specialization.

516
Infectivity, on its own, might lead us to conclude that E. artemiae is a specialist while A. 517 rigaudi is more generalist (also discussed in Lievens et al. subm.). Virulence is difficult to 518 interpret outside the context of spore production, as discussed above, making it a particularly 519 poor proxy for overall specialization. Integrating across all of these life history traits is 520 therefore necessary to properly understand the nature of this host-parasite system, and will  In the mismatched host-parasite combinations, however, up to one third of the exposed hosts 574 were uninfected, and the life histories of these individuals differed clearly from those of 575 control or infected hosts ( This resistance was extremely costly: resistant individuals died much more rapidly than 581 control and infected hosts (Fig. 5). Since there was no detectable compensation through 582 increased fecundity, we must conclude that resistance in these cases is maladaptive. This is 583 intriguing, because A. franciscana and A. parthenogenetica are regularly exposed to their  were less likely to be brooding a clutch when they were infected with A. rigaudi, while in our   Host-parasite combination 2 NA Notes: 1 Survival models were parametric; the best survival distribution was chosen by AICc. 2 A. parthenogenetica exposed to low and high doses of E. artemiae treated 891 separately. 3 Most host growth occurred between days 1 and 30 (Supp. Table 2), so only this period was analyzed further. 4 Offspring could be nauplii or cysts. These two 892 offspring types were not directly comparable: they probably require different amounts of energy to produce, and we allowed mortality to occur before counting nauplii. To 893 account for this, we repeated the tests with nauplii weighted twice, equally, or half as much as cysts, and based our conclusions on the overall pattern. 5 Rate of offspring 894 production = total number of offspring / length of the reproductive period. The length of the reproductive period was the difference between the date of death (or 895 censoring) and the date of sexual maturity. 5 Modeled as clutch size as a function of the elapsed proportion of the reproductive period. The reproductive period started at 896 sexual maturity and ended at death (or censoring). 7 LRS calculated as the total number of offspring produced over the study period. 8 Spore count = the number of spores 897 counted in the fecal sample; we did not transform the spore count to spores/mL (≈ spore count * 700) to avoid skewing the error distribution.

898
Subsets: a Only for females that produced at least 1 clutch. b Analyzed for infected individuals only. c Excluded A. p. exposed to high doses of E. artemiae. d Only for A. p.

932
Lines represent the prediction of the best model, points and vertical bars give the observed means and their 933 95% CIs, calculated over intervals of 10%. Weighing the contributions of nauplii and cysts to the total number 934 of offspring generated qualitatively similar results; the results shown here are for equal weights.  0.00 *Shown for the models that weighted nauplii and cysts equally; giving either offspring type a double weight produces qualitatively equivalent results. †Only two resistant females reproduced, so these results should be interpreted with caution. Figure 1. Spore production and host-to-host transmission success in the four host-parasite combinations. These graphs relate the infection success (percentage of recipients infected) to the spore count in the corresponding spore sample (ln + 1 scale) for A. rigaudi (top) and E. artemiae (bottom). Note that the graphs are divided by recipient species, not donor species (see Methods). Each point represents a recipient group; overlapping points shade to black. Lines represent 2 nd -degree polynomial local regression (LOESS) fittings.

Supplementary Figure 2. Overall fitness measures of A. rigaudi (top) and E. artemiae (bottom) infections.
The asymptotic growth rate (ln + 1 scale) is shown as a function of the lifetime transmission success (ln + 1 scale). The asymptotic growth rate should be maximized during epidemics, while the lifetime transmission success, as an estimator of R0, should be maximized in endemic conditions. The median, first and third quartiles are shown by boxplots on the axes. For A. parthenogenetica infected with E. artemiae, the open circles and boxplot represent the females exposed to a high spore dose. Each point represents an infected host; overlapping points shade to black.

Detailed description of analyses in Table 2 (Methods > Experiment 2: Virulence and transmission > Statistical analyses: virulence & transmission)
First, we analyzed the virulence of infections (effect of the parasite on host survival, growth, reproduction, and overall fitness). This was done separately for A. franciscana and A. parthenogenetica. A. parthenogenetica exposed to low and high doses of E. artemiae were treated as separate treatments. Unless otherwise specified, our analyses proceeded as follows: we included our experimentally manipulated factors in a full regression model, used likelihood ratio tests to test their significance, and where relevant carried out post-hoc comparisons using Dunnett's comparisons with a control (i.e. infected-with-A. rigaudi vs. controls, infected-with-E. artemiae vs. controls). Importantly, we only analyzed virulence once we could be certain of individuals' infection status. To do this, we excluded all individuals that died before day 15 (see Methods > Experiment 2: Virulence and transmission > Experimental design and execution), and only compared infected with control individuals. To make sure that we were not missing important events occurring before this cutoff, we repeated all statistical models for exposed vs. control individuals that died before day 15.
We analyzed host survival using parametric survival models. We established a full fixed-effects model for each host species, then determined the best-fitting parametric distribution (Weibull, exponential, extreme, Gaussian, logistic, lognormal, log-logistic, Rayleigh) using the corrected AIC (Hurvich and Tsai 1989). We then tested the significance of the predictive effects as described above. Finally, we confirmed the fit of the model by performing a goodness-of-fit test (comparing the likelihood of the observed data with the likelihood distribution of simulated datasets based on the model predictions). The full model for A. franciscana included Treatment, Sex, Size class, and all double interactions. The full model for A. parthenogenetica included Treatment, Size class, and their interaction. Origin and Batch were included as frailty components for A. franciscana and A. parthenogenetica, respectively, as they could introduce heterogeneity in mortality rates. Data were right-censored on day 60.
To test the effects of parasite infection on growth, we first checked whether there was significant growth between days 1 & 30 and days 30 & 60 (paired t-tests of the size difference between day 30 & 1 and day 60 & 30). Most growth occurred during the first month (see Results), so we analyzed growth between days 1 & 30 further using linear mixed models. For A. franciscana, we looked at the effects of the fixed effects Sex, Treatment, Size class and all their interactions, with Origin as a random effect. For A. parthenogenetica, the full model included Treatment, Size class and their interaction as fixed effects, and Batch as a random effect.
To analyze (female) reproductive success, we decomposed female reproduction into a) time until sexual maturity, b) the probability of producing a clutch, c) the rate of offspring production, d) the timing of offspring production, and e) the type of offspring produced. All models included Treatment, Size class, and their interaction as fixed effects, and Origin or Batch as random (or frailty) effects. The response variables and statistical models were as follows. a) Time until sexual maturity: the number of days until females became sexually mature, analyzed using parametric survival models. As above, we first determined the best survival distributions to use, then tested the significance of the predictive effects. Females were right-censored in case of death. b) Probability of producing a clutch: a binary variable describing whether a female produced a clutch during the experiment or not, analyzed using generalized linear mixed models with a Bernouilli distribution. c) Rate of offspring production: for females that produced at least one clutch, the total number of offspring divided by the length of the reproductive period. The length of the reproductive period was defined as the difference between the date of death (or censoring) and the date of maturity. The data were analyzed using linear mixed models. Offspring could be nauplii or cysts, and these two offspring types were not directly comparable (they probably require different amounts of energy to produce, and we allowed mortality to occur before counting nauplii). To account for this, we repeated the analyses with nauplii weighted twice, equally, or half as much as cysts, and based our conclusions on the overall pattern. d) Timing of offspring production: for females that produced at least one clutch, the clutch size through time. Clutch size was modelled as a quadratic function of clutch date, with clutch date expressed as the elapsed proportion of the female's reproductive period (e.g. for two females reproducing on the 10 th day of sexual maturity, where one died on the 20 th day and one was censored on the 40 th , the elapsed proportions would be 0.5 and 0.25). Timing was analyzed using generalized linear mixed models with a negative binomial distribution; Individual was included as a random variable to control for pseudoreplication. As in (c), we ran models where nauplii were weighted twice, equally, or half as much as cysts, and based our conclusions on the overall pattern. e) Type of offspring produced: for females that produced at least one clutch, a binomial combination of the number of clutches consisting of nauplii vs. cysts, analyzed using generalized linear mixed models.
As a final virulence measure, we estimated the fitness of (female) hosts. Our fitness proxy was the lifetime reproductive success (LRS), calculated as the total number of offspring produced over the study period. This produced a zero-inflated count distribution, to which we fit negative binomial hurdle models. The full models included Treatment, Size class, and their interaction as fixed effects; random effects (such as Origin and Batch) were not supported by the package. As above, we ran models where nauplii were weighted twice, equally, and half as much as cysts, and based our conclusions on the overall pattern.
Next, we analyzed the parasites' transmission (spore production rate, infectiousness of a single spore, and overall fitness). These analyses were combined for infections in A. franciscana and A. parthenogenetica. Unless otherwise specified, we included our experimentally manipulated factors in a full regression model, and used likelihood ratio tests to test their significance. If relevant, posthoc comparisons were carried out using Tukey comparisons.
To estimate the infectiousness of a single spore, we used the results of the transmission assay. We assumed that the establishment of microsporidian infections follows an independent-action model with birth-death processes. This model assumes that a parasite population grows in the host until it reaches an infective threshold, at which point the infection is considered to be established (Schmid-Hempel 2011 pp. 225-6). In our assay, we considered that an infection was established when we could detect it; in other words, the infective threshold corresponded to the threshold for PCR detection (estimated at ~1 000 spores inside the host's body, unpublished data). In these models, the probability per spore to start an infection, p, is equal to − ln ( ) / where D is the spore dose (Schmid-Hempel 2011 pp. 225-6). In our transmission assay, D can be approximated by the number of spores in the fecal sample taken from the donor at the start of the assay (= spore count transformed to spores/mL, or * 700), divided by 5*8 = 40 (fecal samples accumulated over a 5-day period but we only exposed recipients for one day; the inoculum was shared amongst 8 recipients). We calculated a value of p for every replicate in the transmission assay; p was then analyzed using linear mixed models. The model included Recipient species, Parasite species, and their interaction as fixed effects; an Individual-level random effect was included to control for pseudoreplication (each donor host was used to infect a group of A. franciscana and a group of A. parthenogenetica recipients; some donors were also re-used in the transmission assays on day 30 and 60).
We then tested whether the rate of spore production was dependent on the host-parasite combination. We used Spore count, the number of spores counted in the fecal sample, as the response variable in a generalized linear mixed model with a negative binomial distribution. We did not transform the spore count to spores/mL (≈ spore count * 700) to avoid skewing the error distribution. The fixed effects were Host species, Parasite species, and their interaction; an Individual-level random effect was included to control for pseudoreplication. To avoid comparing apples with oranges, we excluded A. parthenogenetica that had been exposed to 10 000 E. artemiae spores from this model. However, we tested separately whether the rate of spore production differed for A. parthenogenetica infected with different doses of E. artemiae (equivalent model with fixed effect Dose). Spore production analyses were carried out for infected hosts only.
As a final measure of parasite success, we investigated parasite fitness in the different host-parasite combinations. For each established infection (i.e. each infected host), we used two measures of spore production as proxies for parasite fitness. First, we calculated a proxy for the 'lifetime transmission success': we summed the number of spores in the fecal samples taken on days 15, 30, 45 and 60 for each infection, then corrected this cumulative spore count by p, the average infectiousness of a single spore in a given host-parasite combination (as calculated above). Second, we calculated an asymptotic growth rate by computing the dominant eigenvalue of a standard Leslie matrix, , where ni is the number of spores in the fecal sample on day i, p is the average infectiousness of a single spore in that host-parasite combination (as calculated above), and si describes whether the host survived until day i (1) or not (0). While the lifetime transmission success is a measure of the basic reproduction number R0, which describes parasite fitness under stable endemic conditions, the asymptotic growth rate is a measure of the net population growth rate, which describes fitness under epidemic conditions (Frank 1996, Hethcote 2000; we included both measures because either situation can occur in the field. We compared the two measures across host-parasite combinations using non-parametric Kruskal-Wallis tests with Dunn's post hoc testing (R package PMCMR, Pohlert 2014). A. parthenogenetica exposed to low and high spore doses of E. artemiae were treated separately.