Genetic differentiation in pesticide resistance between urban and rural populations of a nontarget freshwater keystone interactor, Daphnia magna

Abstract There is growing evidence that urbanization drives adaptive evolution in response to thermal gradients. One such example is documented in the water flea Daphnia magna. However, organisms residing in urban lentic ecosystems are increasingly exposed to chemical pollutants such as pesticides through run‐off and aerial transportation. The extent to which urbanization drives the evolution of pesticide resistance in aquatic organisms and whether this is impacted by warming and thermal adaptation remains limitedly studied. We performed a common garden rearing experiment using multiple clonal lineages originating from five replicated urban and rural D. magna populations, in which we implemented an acute toxicity test exposing neonates (<24h) to either a solvent control or the organophosphate pesticide chlorpyrifos. Pesticide exposures were performed at two temperatures (20°C vs. 24°C) to test for temperature‐associated differences in urbanization‐driven evolved pesticide resistance. We identified a strong overall effect of pesticide exposure on Daphnia survival probability (−72.8 percentage points). However, urban Daphnia genotypes showed higher survival probabilities compared to rural ones in the presence of chlorpyrifos (+29.7 percentage points). Our experiment did not reveal strong temperature x pesticide or temperature x pesticide x urbanization background effects on survival probability. The here observed evolution of resistance to an organophosphate pesticide is a first indication Daphnia likely also adapts to pesticide pollution in urban areas. Increased pesticide resistance could facilitate their population persistence in urban ponds, and feed back to ecosystem functions, such as top‐down control of algae. In addition, adaptive evolution of nontarget organisms to pest control strategies and occupational pesticide use may modulate how pesticide applications affect genetic and species diversity in urban areas.

Following Stan conventions, the Gamma distribution is here parametrized by the shape and rate (i.e. inverse scale) parameters.
We performed prior sensitivity analyses to assess the sensitivity of the posterior inferences to the prior specification. We consider three main scenarios (A, B and C) for the sensitivity analyses. In scenario A, we alter each (group of) parameter(s) separately, in a sequential fashion: in subscenario A1, in A2 and in A3. For each (sub)scenario, we consider a strongly regularizing prior, where the scale parameters are divided by 5: ~StudentT ( In scenario B, we specifically investigate prior choice for the number of degrees of freedom of the Student's t-distributed clone-level random effects, where we evaluate an alternative prior choice that places more prior mass towards lower values of : and an alternative prior that is wider than the original choice: ~Gamma(3,0.05).
In scenario C, we simultaneously alter the prior choice for all parameters , and , except for . Here too, we consider the same strongly regularizing priors, regularizing priors, vaguer priors and strongly vaguer priors as outlined for scenarios A1-A3.
For each scenario (A, B1, B2, B3 and C), we refit the model with the alternative prior specifications and plot and compare the obtained posterior densities with these of the original model. We use 4 chains of 2,000 iterations each (of which 1,000 are discarded as warmup) instead of 10,000, as this offers a major speedup without affecting MCMC convergence (due to excellent mixing). Specifically, we monitor changes with respect to the posterior densities of , , , and the log probability (up to an additive constant).
The prior sensitivity analyses do not reveal evidence that the presented posterior inference is particularly sensitive to the prior specification ( Figures S6-S10). Alternative prior 7 specifications for the regression parameters have the strongest influence on posterior inference, though differences are relatively modest. The two key effects discussed in the manuscript (main effect of pesticide exposure and its interaction with an urban origin) maintain a high probability of a respectively negative and positive effect under all evaluated scenarios, even under a strongly regularizing prior with unit variance. In some cases, the strongly vaguer prior specification causes the Hamiltonian Monte Carlo procedure to produce a few divergent transition: too vague priors are known to frustrate Hamiltonian Monte Carlo.

Posterior predictive checking
We asses goodness-of-fit by means of a simple graphical posterior predictive check. For this, we sample 20,000 draws (4 chains x 5,000 post-warmup draws) from the posterior predictive distribution, as part of the MCMC procedure. We then compare the distribution of frequencies of the number of surviving individuals with the observed frequency of the number of surviving individuals. This graphical posterior predictive check does not reveal strong deviations of the posterior predictive distribution compared to the observed data ( Figure S11). However, the model does underpredict the number of replicates with low and high survival to a moderate extent (mainly for instances with 0 and 5 surviving individuals).
Hence, there seems to be some overdispersion with respect to the considered binomial distribution. In order to ascertain that our findings are robust against the presence of overdispersion, we developed a second version of the model that includes an observation-level random effect (which has a similar effect as using a beta-binomial distribution). Stan code for this model is also available online ("Overdispersion_model.stan" on https://github.com/mfajgenblat/brglmm-chlorpyrifos) and can completely be run from the original R-script after changing the model file name argument in the stan()-function.
Graphical posterior predictive checks indicate that his second model is able to accommodate overdispersion ( Figure S12). Compared to the original model, this extended model yields close to identical inferences ( Figure S13).