Notice: Wiley Online Library will be unavailable on Saturday 27th February from 09:00-14:00 GMT / 04:00-09:00 EST / 17:00-22:00 SGT for essential maintenance. Apologies for the inconvenience.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
Ecological effect assessment is the part of the ecological risk assessment in which one aims to set a safe ecological threshold concentration for the chemical of interest. Below this safe ecological threshold, no effects on ecosystem structure and function are expected. Among the most realistic approaches to assess the ecological effects of chemicals are ecosystem-level experiments using micro- or mesocosms in which measures of ecosystem structure or function are monitored over time and at different chemical concentrations. These approaches generally are considered to be environmentally relevant and realistic because they allow researchers to determine endpoints at higher levels of biological organization and assess intra- and interspecific interactions and indirect effects 1. However, non-trivial interpretation of test results and the time and resources required make these approaches less attractive for routine use. Therefore, models have been proposed as tools that use results from less resource-demanding, single-species toxicity tests to predict safe ecological threshold concentrations in ecosystems. Models that simultaneously account for the sensitivity of multiple species include toxicity-extrapolation models based on species sensitivity distributions (SSDs) and food web models. The SSD is a probability distribution that is fitted to a set of toxicity thresholds from individual test species 2. Food web models simulate population dynamics in ecosystems exposed to chemicals by describing predator–prey relationships, metabolic costs, and environmental forcing and by imposing chemical effects on the modeled species according to the toxicity thresholds of individual test species 3, 4.
In the literature, the use of SSDs and food web models to assess the ecological effects associated with the presence of chemicals in the environment has been evaluated. For SSDs, the influence of the chosen statistical distribution, the quality and quantity of the used toxicity data, and the influence of the included species and trophic levels on the predicted ecological threshold have been examined 2, 5–8. It should be noted, though, that most of the research on SSD modeling has focused on pesticide applications, most notably insecticides 5, 9 and herbicides 10 and to a lesser extent fungicides 11. Similar exercises for other organic chemicals 12 or metals are less abundant 13, 14. Compared to SSD modeling, food web models have been applied less frequently in the literature on risk assessment 15, and only a few studies have done so 16. In a regulatory context in Europe, the use of food web models is non-existent. Despite the demonstrated capacity of such models to obtain protective ecosystem-level thresholds 4, 15–17, none of the risk assessments performed under the EU regulation 793/93/EC has used food web modeling as a basis for deriving predicted no effect concentrations (PNEC).
In the present study, we re-evaluate the environmental risk assessments of 78 existing chemicals performed under EU regulation 793/93/effect concentration (EC) between 1996 and 2009, using two types of effect models: SSDs and food webs. The present exercise has three goals. First, the accuracy of the two effect modeling approaches was examined for a small subsample of the evaluated chemicals (n = 7) by comparing the predicted threshold concentrations to threshold concentrations observed in micro- or mesocosm experiments. Second, the uncertainty on this predicted threshold was assessed and compared between the two modeling approaches. Third, the safe ecological threshold concentrations predicted by effect modeling were compared to the final PNECs for the aquatic environment as they are listed in the risk assessment reports.
MATERIALS AND METHODS
Risk assessment reports
Risk assessment reports of existing chemicals evaluated under regulation 793/93/EC were obtained from the European Chemicals Bureau (ECB) website (http://ecb.jrc.ec.europa.eu/existing-chemicals/) and the included single-species (no observed) acute and chronic effect concentrations (hereafter collectively referred to as ECs for simplicity, although no-observed-effect concentrations (NOECs) were also included) were extracted. Only the reported bounded ECs were used. Therefore, ECs that were defined only in a qualitative way, for example, by an inequality (i.e., <x µg L−1) or using ambiguous terminology (e.g., ∼x µg L−1) were discarded. Furthermore, only experimentally measured ECs were retained; that is, estimates from quantitative structure activity relationships were not considered. In addition, ECs for metabolites of the chemicals were not included. In cases where both marine and freshwater effect concentrations were available for a chemical (e.g., naphthalene), the ECs for both environments were treated separately in two different risk assessments. Chemicals for which only a single EC was available were not considered further. Categorizing an EC as acute or chronic was determined based on the exposure duration of the toxicity test 2. In the present study, acute toxicity tests were characterized by an exposure duration of <12 h for algae and ≤4 d for invertebrates and vertebrates. In chronic toxicity tests with invertebrates and vertebrates, effects are studied over prolonged periods of exposure, that is, ≥12 h for algae and >4 d for invertebrates and vertebrates. An overview of the 78 chemicals considered in the present study is given in Supplemental Data, Table S1. The full database of ECs can be obtained on request. A summary on the data availability per chemical is shown in Supplemental Data, Figure S1.
For each chemical, 42 (=2 × 7 × 3) SSDs were constructed that differed in the type of statistical distribution (log normal; log logistic); the included trophic levels (only primary producers; only invertebrates; only vertebrates; all three combinations of two of the three trophic levels; all three trophic levels); and the type of ECs selected (all acute ECs; all chronic ECs; only chronic EC10s and chronic NOECs). When several EC values based on the same toxicological endpoint were available for one species, the geometric mean was used, which resulted in a species mean EC. When several EC values based on different toxicological endpoints were available for one species, the lowest value was selected. The lowest value was determined based on the geometric mean if more than one value for the same endpoint was available. Parametric SSD models were fitted using maximum likelihood estimation using the function fitdistr in R 18. The significance of the fits was assessed using a chi-square test with p = 0.05. Confidence intervals of the fifth percentile (i.e., the hazardous concentration for 5% of the species; HC5) were obtained from 10,000 independent Monte Carlo runs. In every run, a different value was assigned to the parameters of the fitted distribution (i.e., mean and standard deviation for the log normal distribution; location and slope for log logistic distribution). The different values for these parameters were sampled randomly from their 90% confidence intervals, as they resulted from maximum likelihood estimation. Throughout the present study, only SSDs that were found to fit the data significantly based on chi-square testing were considered.
Food web model description
A Rosenzweig-MacArthur food web model 19 was constructed in R 18. The model considered in the present exercise included 20 species: 12 primary producers, six invertebrates (consuming primary producers), and two vertebrates (consuming invertebrates). As such, the trophic composition of the proposed food web was realistic, because the ratio of the number of primary producers, invertebrate species, and vertebrate species was within the range found in natural systems 20. The food web model was equipped with logistic concentration–response functions to describe the direct toxic effects of a chemical concentration, c (input variable), on the reproduction and survival of the invertebrates and vertebrates and on the carrying capacity and photosynthesis rate of the primary producers 16. Parameters of these functions were the median lethal concentration (LC50) and the reproduction-EC50 for animals and the EC50 for growth rate of primary producers. Effects on the photosynthesis rate and the (in)vertebrates' mortality rate(s) were implemented as done previously 4. Effects on the carrying capacity were assumed to be equal to the effects on the photosynthesis rate, based on a recent meta-analysis 21. The food web model does not explicitly include reproduction of invertebrates and vertebrates in its formulation. Instead, it calculates net biomass production as the difference between biomass gain (assimilation) and loss (mortality) for these populations. An x% effect of a chemical on reproduction was therefore modeled as an x% reduction in the gain process. All model equations are shown in Figure 1. Parameters and variables are described in Table 1.
Table 1. Parameters (min, max) and variables used in the food web modela
All values for primary producers, invertebrates and other parameters are taken from Dakos et al. 22. Values for vertebrates are as reported in De Laender et al. 4 and Hendriks et al. 38. Slope values for concentration response functions were taken from Smit et al. 39.
Indicates that the parameter is chemical/data specific.
Maximum growth rate of primary producer i (d−1)
Competition coefficient among primary producers i and j (–)
Carrying capacity for primary producer i (mg L−1)
50% effect concentration for phytoplankton growth (µg L−1)
Slope of concentration-response functions for vertebrate l
Background mortality rate of vertebrate l (d−1)
Immigration rate (mg L−1 d−1)
Amplitude of seasonal forcing (–)
Seasonal forcing (–)
Biomass of primary producer i (mg L−1)
Biomass of invertebrate k (mg L−1)
Biomass of vertebrate l (mg L−1)
Mortality rate of invertebrate k (d−1)
Mortality rate of vertebrate l (d−1)
Chemical concentration (µg L−1)
Chemical effect on growth rate and carrying capacity of primary producer I (–)
Chemical effect on reproduction of invertebrate k (–)
Chemical effect on reproduction of vertebrate l (–)
Food web model parameterization
Seasonal forcing of population growth rates and carrying capacity were parameterized so that the modeled community was representative for a mesotrophic, temperate environment 22. Seasonal forcing is a sinusoidal function, defined by a period and amplitude. Functional diversity within one trophic level was incorporated by allowing functional traits (e.g., growth rates of phytoplankton species) to vary within one trophic level 23. To create these functionally diverse communities, random values were iteratively sampled for the parameters from their corresponding ranges, and the model was run for 10 years. After every model run, total biomass was calculated per trophic level. Next, the decrease of total biomass per trophic level (phytoplankton, invertebrates, and vertebrates) was calculated. If the decrease was a factor 2 to 100 per trophic level 24, biomass ratios between trophic levels were categorized as realistic. When this was the case, iterations stopped and the resulting community was used for all the food web modeling described in the present study. In Supplemental Data, Figure S2, the resulting community dynamics are shown for 30 d. A list of the parameter values used and the interpretation of model variables are listed in Table 1.
Food web model simulations
Details regarding how the ecological effects were calculated and how the HC5 was obtained using the food web model are given in the Supplemental Data (Food web modeling: HC5 derivation). Briefly, the effects of a given concentration of a given chemical X on the population sizes (biomasses) of all species in the food web were calculated by comparing the average control biomasses of the 20 species to those in the exposure situation. The latter were simulated by setting the concentration c (Fig. 1) equal to that concentration and setting the LC50s and EC50s of the food web model to those ECs listed in the risk assessment report for chemical X. This parameterization was done by means of bootstrapping with replacement to account for uncertainty in the toxicity data. This parameterization required converting the effect levels of the ECs in the risk assessment reports to those endpoints and effects levels (50%) needed in the food web model. The HC5 was calculated by iteratively changing c until exactly 5% of the species was affected; in other words, 1 of the 20 species in the food web. A species was categorized as affected if its mean biomass over time was 20% lower than that in the control 25. The bootstrapping procedure also allowed us to calculate the uncertainty range of the HC5; that is, its 90% confidence interval.
Evaluating model simulations
Accuracy of the toxicity–extrapolation and food web models was examined by comparing predicted HC5-50s to community-level NOECs observed in multi-species experiments (Supplemental Data, Table S2). If multiple SSDs fitted the data for one chemical, the following approach was adopted: If the SSD including all species significantly fitted the data, the corresponding HC5 was retained and compared to the multispecies experiment. If this was not the case, the lowest HC5 of all SSDs significantly fitting the data was retained for comparison with the multispecies data. Only community NOECs for endpoints comparable to the one considered by the food web model, that is, biomass or abundance of populations, were considered in the present exercise. It is important to note that community NOECs were available for seven chemicals only (nonylphenol, phenol, trichloroethylene, bisphenol A, Cd, Cu, Zn); that is, for approximately 10% of the total number of chemicals assessed. For Zn and Cu, six and three micro- and mesocosm studies were found from which a community NOEC could be derived, respectively. Because of the known influence of physicochemical water characteristics on metal bioavailability and toxicity, the available single-species ECs for copper and zinc (all chronic) were normalized to the water composition prevailing in these micro- and mesocosm studies using chronic biotic ligand models (BLMs) for Cu 26, 27 and Zn 28–31, respectively. For both metals, chronic BLMs have been developed and validated for a limited number of species, usually a cladoceran (Daphnia magna), a fish (Oncorynchus mykiss or Pimephales promelas) and an alga (Pseudokirchneriella subcapitata or Chlorella vulgaris). However, the available toxicity data for these metals include species other than just these three for which BLMs have been developed. We assumed that all invertebrate toxicity data could be normalized with the D. magna BLM, all vertebrate (fish) NOEC values with the O. mykiss BLM, and all algae NOEC values with the P. subcapitata BLM. Based on this normalisation procedure, specific HC5-50s were derived using the SSD and food web models as described above for three (for copper) and six (for zinc) micro- or mesocosms.
The HC5-50s obtained using SSD and food web modeling were compared to the final PNECs aquatic as they are listed in the risk assessment reports. In the majority of cases (>90%), these PNECs were derived using so-called extrapolation or assessment factors, that is, the lowest toxicity datum was divided by an arbitrary number that increases as data availability decreases.
RESULTS AND DISCUSSION
Fit of SSD models
The fit of the SSDs to the toxicity data was significant for 0 to 35% of the chemicals considered, depending on the type of toxicity data used (acute vs chronic) and the SSD's trophic composition (Supplemental Data, Fig. S3). The two types of SSD models used (log normal vs log logistic) were equally successful in fitting the toxicity data, which agrees with earlier findings 32. Species sensitivity distributions based on acute toxicity data only were most successful in fitting these data when using data from invertebrates and vertebrates. Species sensitivity distributions based on chronic data fitted the data from all trophic levels for 35% of the chemicals, which confirms the results from other studies 8. In general, SSDs including the toxicity data of primary producers alone performed worst. This can be understood by the low availability of toxicity data for primary producers (Supplemental Data, Fig. S1). In general, the number of data points influenced the significance of SSD fits. Indeed, 10 to 15 data points were necessary to obtain a significant fit with a log normal SSDs if NOECs or EC10s from all trophic levels were included (Supplemental Data, Fig. S4; light blue symbols). When based on all chronic ECs, fewer data points were sufficient, in general, to obtain a significant fit (Supplemental Data, Fig. S4; dark blue symbols). This finding agrees with previous studies on the minimum number of data points that are needed to obtain reliable HC5-50 estimates, which range from 10 6 to 30 8.
SSD and food web model predictions
The fraction of affected species as predicted by the food web models was always larger than or equal to the corresponding prediction by the SSDs. This is illustrated for an organic chemical (bisphenol A) and a metal (Cu) in Supplemental Data, Figure S5. For bisphenol A, the median food web predictions suggested 5% of the species would be affected at concentrations from 2 µg/L up to 20 µg/L, whereas the SSD predicted a monotonous increase in the affected fraction over the entire concentration range. Predicted effects of Cu (assuming the water chemistry of Roussel 33) were comparable between the two approaches. In the highest concentration range, however, both approaches diverged in their assessment of ecological effects. Possibly, such differences can be attributed to the sensitivity of phytoplankton, with phytoplankton being less sensitive than the higher trophic levels causing deviations between the SSD and the food web model. Indeed, for Cu (small differences between SSD and food web model), the mean NOEC of phytoplankton (16.8 µg/L) was lower than the mean NOEC for invertebrates (31.7 µg/L) and the mean NOEC for vertebrates (50 µg/L). Conversely, the average NOEC of phytoplankton for bisphenol A (large differences between SSD and food web model) was four times higher than that of invertebrates and vertebrates. The same observation was made for 3,4-dichloroaniline—large differences between the SSD and the food web model (results not shown) corresponded to phytoplankton being less sensitive than invertebrates and vertebrates.
SSD and food web model HC5-50s
Median hazardous concentrations for 5% of the species (HC5-50) were estimated by the food web model and by SSD modeling. The food web model produced HC5-50s that were better predictors of the measured community NOECs than the HC5-50s produced by log normal or log logistic SSDs (Fig. 2 and Supplemental Data, Fig. S6). However, it should be acknowledged that model predictions could only be compared to data for seven chemicals (nonylphenol, phenol, trichloroethylene, bisphenol A, Cd, Cu, Zn). The HC5-50s predicted by the food web model varied mostly between 0.5 and 2 times the observed median community-level NOECs. Accuracy did not change with data availability, quantified as the number of species for which data were available. Species sensitivity distribution models based on acute data produced HC5-50s that were either accurate or higher than the observed community-level NOECs. The HC5-50s from the SSDs based on chronic data were up to 50 times lower than the observed community-level NOECs. In line with previous exercises 6, 7, 9, the influence of the chosen distribution type on the HC5-50 was very small (< factor 1.5).
The HC5s that were predicted by the food web model were sensitive to the effect size chosen on a population level. By default, following earlier exercises 4, 15, a population was categorized as affected when its size, quantified as biomass density, was reduced by at least 20%. However, the food web model allows considering alternative effect sizes as well, if a particular risk assessment should so require. A sensitivity analysis performed for one of the metals (Cu, assuming the water chemistry of Roussel 33) and one of the organic chemicals (bisphenol A) revealed a 4-fold (Cu) and 3-fold (bisphenol A) change between the modeled HC5-50 when changing population-level effects from 5 to 40% (Supplemental Data, Fig. S7).
More measured community-level NOECs were available for metals than for organic chemicals. Although data for only three metals were available, multiple community-level studies, and thus NOECs, were available for each of these three metals (two for cadmium, three for copper, six for zinc). These studies differed in their bioavailability-determining water characteristics such as pH and DOC concentrations and therefore in their community-level NOEC. These differences were predicted well by both modeling approaches (Supplemental Data, Fig. S8 in the Supplemental Data illustrates the results for the food web model).
The finding that SSDs based on chronic toxicity data produce HC5-50s that are lower than observed ecological thresholds 10, 12, 14, 34 is often suggested to be the result of the lack of random species selection during SSD construction. Indeed, taxa used in single-species toxicity tests are often hypothesised to be among the most sensitive taxa found in the field 5, 14. In the present study, we have used the same toxicity data for SSD modeling as for food web modeling. Most of the HC5-50s from SSDs based on chronic data were 2 to 50 times lower than the observed community NOECs, whereas the food web model produced HC5-50s that were closer to the observed community NOECs. This suggests that the observed differences between chronic SSD-based HC5-50 values and observed community NOECs do not necessarily imply inherent sensitivity differences between species used in a single species toxicity test and the species found in the (semi-) natural systems evaluated in micro- and mesocosm experiments. Instead, our results suggest that differences between predicted and observed community responses to chemical exposure may indicate limitations of a given methodology to predict community responses rather than inherent differences in species sensitivity.
The 90% confidence intervals of the estimated HC5s were comparable for both modeling approaches for sample sizes >10 (Fig. 3). For sample sizes <10, no chronic SSDs significantly fitted the toxicity data and therefore no confidence interval was obtained. Note that we only consider the SSDs with chronic data, as these most often fitted the toxicity data significantly. For SSDs based on chronic ECs, the 90% confidence intervals of the estimated HC5 seemed to decrease slightly with increasing data availability, but this decrease was not significant (linear model; p > 0.05). The 90% confidence intervals of the HC5s predicted by food web models were, on average, smaller than those obtained by SSD modeling and were also not significantly related to the amount of toxicity data available (linear model; p > 0.05).
HC5-50 obtained with SSDs and food web modeling: A comparison with the PNEC
For all chemicals, the HC5-50 derived using the two modeling approaches were higher than the final PNEC for the aquatic environment as reported in the risk assessment reports (Fig. 4). We should note, however, that initially the HC5-50 for dibutylphthalate we calculated with the food web model was lower than the PNEC for this substance. This is because we used the 96 h EC50 for Gymnodium breve of 0.0034 mg/L, although it was judged unreliable—and therefore not considered—in the risk assessment report due to reproducibility issues. Therefore, we discarded this value and recalculated the HC5-50 with the food web model as 140 µg/L. In general, differences between the PNEC and the HC5-50 were larger for organic chemicals (factor 10 to 100) than for metals (factor 2 to 5). The most frequently adopted approach in deriving PNEC has been to divide the lowest single species ECx by an assessment factor to account for the uncertainty when extrapolating from the individual single-species level to the ecosystem level. In general, these assessment factors decrease as the amount of chronic data increases 35. Therefore, we expected that the difference between the PNEC and the HC5-50 would decrease as the amount of available toxicity data increases. For chemicals with few data available, we expected larger differences between the PNEC and the HC5-50 than for well-studied chemicals. However, the difference between the PNEC and the HC5-50 did not decrease as data availability increased. Instead, these differences varied by several orders of magnitude for chemicals with comparable data availability (e.g., factors difference between 5 and 800 when <10 chronic (NO)ECxs available; Fig. 5). A possible explanation for this is that the HC5-50s we report in the present paper were derived using all toxicity data (food web model) or a subset of it (SSD models, e.g., using only acute or chronic data). Thus, the HC5-50s obtained using effect modeling summarized the available toxicological knowledge on the chemicals for which risk has to be assessed. In contrast, the PNECs that are derived using the assessment factor approach are driven by the lowest (NO)ECx and neglect information on the toxicity toward slightly less sensitive and tolerant species, which reflects the basic assumptions on which the technical guidance document for risk assessment (TGD) is based. In other words, the toxicity data on the most sensitive species alone is adequate to assess the sensitivity of ecosystems 35.
The comparison between the reported PNEC in the risk assessment reports and the HC5-50s predicted by the two modeling approaches reflects the principle of tiering well, although not as included in the TGD documentation. Being most often obtained using the so-called assessment factors, the regulatory PNECs can be considered results from effect assessments made at the lowest tier. Because the aim of lower tier approaches is to assess effects in a conservative (i.e., worst case) manner, they are expected to produce ecological thresholds that are lower than those produced by higher-tier methods such as SSDs and food web models. For all chemicals of which effects were assessed, this pattern was found, because the thresholds the models produced were higher than those given by the risk assessment reports.
The present contribution demonstrates that statistical and mechanistic effect models are capable of predicting ecological thresholds that are protective for the seven chemicals for which the models could be tested. Previous contributions have demonstrated the predictive capacity of such models for other chemicals as well when comparing their predictions to the results from micro- and mesocosm studies 4, 5, 10, 12, 13, 36, 37. However, in the context of effect assessments, effect models are typically used to obtain point estimates of concentrations at which a certain ecological effect (or the absence of effect) is expected. In the present study, we adopted the same approach because we wanted to be able to compare model predictions with the available data, which were essentially point estimates of ecological (no observed) effect. It should be noted, however, that effects models can be used in a much more informative way by considering the whole concentration–response relation they predict. Although the thresholds (HC5-50s) that were predicted by the models were always higher than the PNECs listed in the risk assessment reports, it was impossible to explore the influence of the exposure concentration on the ecological effect observed in the micro- and mesocosm experiments, because these experiments rarely provide the full concentration-response relation. When such data becomes available, more detailed model evaluations will be possible by comparing the predicted and observed concentration–response. Tested models would be able to (partly) replace resource-demanding mesocosm studies in the future. Based on predictions of these models, and on the associated uncertainty, more accurate and ecologically relevant risk assessments are possible.
Tables S1 and S2. (195 KB PDF).
F. De Laender is a postdoctoral research fellow from the Research Foundation— Flanders.