Bioassays
Salmon-louse bioassays were performed by the BC Centre for Aquatic Health Sciences (CAHS) as described in Saksida et al.10. Briefly, motile (i.e., pre-adult and adult) L. salmonis were collected from 11 salmon farms in the Broughton Archipelago (BA) between 2010 and 2021 and transported to CAHS in Campbell River, BC. Within 18 h of collection, healthy lice were separated by sex and randomly placed into petri dishes each containing approximately 10 lice (mean ± SD = 9.6 ± 1.1) and subjected to one of six EMB concentrations (either 0, 31.3, 62.5, 125, 250, and 500 ppb or 0, 62.5, 125, 250, 500, and 1000 ppb, depending on suspected variation in EMB sensitivity11). Each collection corresponded to one bioassay, and each bioassay contained roughly four replicates for each sex (4.0 ± 1.3 for females and 3.6 ± 0.9 for males). After 24 h of EMB exposure, lice were classified as alive if they could swim and attach to the petri dish, or moribund/dead otherwise. Lice were kept at 10 °C throughout the process. In total, 34 bioassays were conducted from 11 farms between October 2010 and November 2021.
We analysed the proportion of lice that survived exposure to EMB, using standard statistical descriptions that accounted for within-assay dependencies (generalized linear mixed models (GLMMs) with logit link functions, fitted separately to the data from each bioassay). The models included fixed effects for EMB concentration, sex, and the interaction between the two, as well as a random intercept for petri dish. For each analysis, we centered concentration values and scaled them by one standard deviation. We used the GLMM fits to calculate the effective concentrations at which 50% of the lice survived (EC50) in each bioassay. The GLMM for one bioassay produced a singular fit because there was not enough variation in the female survival data to warrant the random-effects structure. We retained the EC50 values resulting from this singular fit because re-fitting without the random intercept yielded identical EC50 values, and removing the entire bioassay from the overall dataset did not qualitatively affect the subsequent analysis.
To assess whether the sensitivity of salmon lice to EMB has decreased over time, we fitted a set of five standard GLMs with gamma error distributions and log link functions to the maximum-likelihood EC50 estimates. Each of these five models included binary effects for sex and for whether the farm’s stock had previously been treated, since both affect EMB sensitivity in lice10. The first model included only these two effects and served as a null model that assumed lice did not evolve EMB resistance over time. The second model added a fixed effect for time (i.e., the number of days since January 1, 2010), while the third model included an interaction between time and sex. The fourth and fifth models were identical to the second and third, but with a quadratic effect for time, to account for possible first-order nonlinearity. We were unable to add an effect for farm due to small sample sizes. We performed model selection using the Akaike Information Criterion penalized for small sample sizes AICc25, treating AICc differences of less than two as being indistinguishable in terms of statistical support and selecting the least complex model when that was the case26. The ΔAICc values for the EC50 models were 48.1, 6.1, 4.9, 0, 1.75, respectively.
Field efficacy
We used relative salmon-louse counts after EMB treatment (i.e., the post-treatment count divided by the pre-treatment count) as our measure of EMB field resistance between 2010 and 2021 (higher relative counts imply lower treatment efficacy). We defined “pre-treatment” as one month prior to treatment and “post-treatment” as three months after treatment (roughly when one would expect to find the lowest counts in louse populations previously unexposed to EMB), as in Saksida et al.10. We excluded EMB treatments for which an additional, non-EMB treatment was performed within the following three months. In total, there were 73 EMB treatments for which we were able to calculate relative post-treatment counts.
Salmon-louse counts were performed by farm staff as described by Godwin et al.27. In short, salmon-louse counts were usually performed at least one per month by capturing 20 stocked fish in each of three net pens using a box seine net, then placing the fish in an anesthetic bath of tricaine methanesulfonate (TMS, or MS-222) and assessing the fish for motile (i.e., pre-adult and adult) L. salmonis by eye.
The treatment dataset included the date and type of every treatment that has been performed on a BA farm (i.e., not just the 11 farms with bioassay data). In total, 88 EMB treatments were conducted between 2010 and 2021, of which we were able to calculate relative post-treatment counts for 73 because some months lacked counts or had a non-EMB treatment performed within the following three months. An additional 22 non-EMB treatments (e.g., freshwater and hydrogen baths) were performed, all since the beginning of 2019, but we excluded these data from our analysis.
To determine whether field efficacy of EMB treatments has decreased over time, we used GLM-based “hurdle models”—standard statistical descriptions used to accommodate an over-abundance of zeroes in data being analysed. A hurdle model uses two components—one model for whether a count is nonzero and another for the value of the nonzero count—to predict overall mean count. To this end, we fitted three binomial GLMs paired with three gamma GLMs to the relative-count data, each of the paired models being structurally identical in terms of predictors. All of these submodels included a binary fixed effect for previous treatment, as in the EC50 models. The null pair of submodels included no additional terms, the second pair of submodels included a fixed effect for time (i.e., the number of days since January 1, 2010), and the third pair of submodels included a quadratic effect of time (again, to account for possible first-order deviations nonlinearity). We were unable to add an effect for farm due to small sample sizes. We performed model selection of the hurdle models, again using the Akaike Information Criterion penalized for small sample sizes. The ΔAICc values for the three hurdle models were 39.6, 18.3, and 0, respectively. We performed our analyses in R 3.6.028, using the lme4 package29.
Source: Ecology - nature.com
