Estimating species sensitivity distributions on the basis of readily obtainable descriptors and toxicity data for three species of algae, crustaceans, and fish

Introduction

In the ecological risk assessment of chemicals, the use of species sensitivity distributions (SSDs) is fundamental to deriving “safe” concentrations such as the predicted no effect concentration (PNEC) and water quality benchmarks, as well as the potentially affected percentages of species (Vaal et al., 1997; Posthuma, Suter & Traas, 2002; Iwasaki et al., 2015; Belanger et al., 2017; Carr & Belanger, 2019; Posthuma et al., 2019; Sorgog & Kamo, 2019). SSDs can be derived on the basis of acute or chronic toxicity data, or both. To derive an SSD for a given chemical, a set of toxicity data is fitted to statistical distributions such as a log-normal or log-logistic distribution. Therefore, toxicity results for multiple species are required, and the minimum sample sizes (i.e., the minimum number of species required) typically vary between five and 10, depending on the regulatory jurisdiction (Belanger et al., 2017). This data-demanding feature is a critical barrier to the application of SSD to chemicals for which there are few, or no, toxicity data (Hoondert et al., 2019; Hiki & Iwasaki, 2020).

To address this data-demand issue, the development of statistical models with physicochemical descriptors to estimate the mean and standard deviation (SD) of the logarithms of log-normally distributed SSDs has recently been proposed (Hoondert et al., 2019). Development of separate models for the mean and SD of the logarithms of SSDs rather than models that estimate the hazardous concentration for X% of species (HC_x; typically, HC₅) based on an SSD is valuable for two different reasons. First, both of these parameters are required to estimate the fraction of species affected by a given concentration of a toxic substance as well as the concentration that is hazardous for a given percent of the species (Posthuma, Suter & Traas, 2002). Second, the mean and SD have inherently different statistical properties (e.g., the SD of the logarithms of the SSD is related to inter-species differences in sensitivity). Although the results are preliminary, the coefficients of determination (R²) of models developed for the SDs of the logarithms of the SSDs were less than 0.4 (Hoondert et al., 2019). However, the parameters that characterize SSDs can be more robust than the concentrations that are acutely toxic to individual species (Iwasaki et al., 2015). A modeling approach aimed at characterizing SSDs therefore seemed worth pursuing.

The major concept that we tested here was to make use of the mean and SD of toxicity values for three biological species chosen (randomly) from three biological groups (algae, crustaceans, and fish), in addition to readily obtainable descriptors, to improve the model prediction of SSDs. This idea is similar to the application of QAAR (quantitative activity–activity relationship) and QSAAR (quantitative structure–activity–activity relationship) models (Furuhama, Hayashi & Yamamoto, 2019). Acquiring at least one set of toxicity data from each of algae, crustaceans, and fish may not be feasible for many chemicals, but these biological groups are often the most representative in terms of data availability (Aurisano et al., 2019). Given the difficulties in predicting SSD parameters on the basis of readily obtainable descriptors alone (Hoondert et al., 2019), an approach that lies in between prediction based on “only descriptor-based” models and prediction based only on experimentally obtained toxicity data could be promising.

Here, acute toxicity data were used to estimate SSDs because of the paucity of quality-assured acute toxicity data. Our specific aims were threefold (Fig. 1). We first aimed to develop and test linear regression models with readily obtainable descriptors, such as the log₁₀-transformed octanol–water partition coefficient (log K_OW), to predict the means and SDs of acute log-normal SSDs. By doing this, we evaluated the previous conclusions of Hoondert et al. (2019) about model performance. Second, we aimed to develop linear regression models with predictors of means and SDs estimated on the basis of toxicity values for three species chosen from three biological groups (algae, crustaceans, and fish). Finally, we aimed to develop and test linear regression models with predictors of means and SDs based on the three species’ toxicity data, as well as the descriptors, to predict the means and SDs of SSDs. Although several studies have derived SSDs based on three species’ toxicity data and have assessed the associated uncertainties of HC₅₀ and HC₅, particularly in the field of life-cycle impact assessment (Van Zelm et al., 2007; Golsteijn et al., 2012; Douziech et al., 2020), to our knowledge no attempt has been made to develop and test the models by the combined use of the two different types of predictors (i.e., descriptors and three species’ mean and SD).

Materials and Methods

Results

Predictors included in the best model developed with descriptors alone for SSD mean were Class 4, log K_OW, and Class 3, all of which had higher (negative) full-model-averaged standardized parameter estimates than the other variables (<−0.19; Table 2). Similarly, the best model developed with descriptors alone for SSD SD included only Class 4. However, along with Class 4, Log VP and MW had higher full-model-averaged standardized parameter estimates than the other variables (Table 2). Although the RMSE and RMSE_ext values of these best models were both below approximately 0.8 for the SSD mean and below approximately 0.2 for the SD (i.e., both within a factor of 10 of the values based only on measured toxicity data), the r² values were low (below approximately 0.6; Table 3 and Fig. 2).

Incorporating the means or SDs of the toxicity values for three species markedly improved the performances of the models (e.g., see changes in r² values in Table 3; Fig. 2). The best models for both descriptors and 3-species-based mean/SD were (see the Supplementary File for more details):

(2)${\displaystyle \mathrm{SSD mean}=0.07+1.02\times 3\text{-species}\mathrm{mean}-0.27\times \mathrm{Class}4}$ (3)${\displaystyle \mathrm{SSD SD}=0.39+0.48\times 3\text{-species SD}+0.30\times \mathrm{Class}4+0.02\times \mathrm{log V\; P}.}$

Our finding that the differences in AICc values between models with descriptors alone and those with the 3-species mean or SD were over 20 also indicated the substantial model improvements with the latter (Table 3; also see Burnham, Anderson & Huyvaert, 2011; Iwasaki & Hayashi, 2020). Particularly for the models for SSD mean, inclusion of only the 3-species mean as a predictor increased the r² values from 0.62 to 0.96, and this was also evident from the prominently high full-model-averaged standardized parameter estimates of this predictor (i.e., 0.94; Table 2). Inclusion of only the 3-species SD also improved the model fit in the prediction of SSD SD (e.g., r² increased from 0.49 to 0.70; Table 3), and model performance was further improved by including two descriptors (Log VP and Class 4; Tables 2 and 3). The best model for predicting SSD mean had RMSE and RMSE_ext values of <0.5, and that for predicting SSD SD had RMSE and RMSE_ext values of <0.2 (Table 3). Moreover, when we selected and evaluated the best models by using five randomly chosen bootstrapped samples, the statistics for model evaluation usually changed little (particularly, the coefficients of variation in the three statistics for the internal validation (r², RMSE, and q²) were <6%), although the best models sometimes had one more, or one less, predictor compared with the best models developed above (see Tables S2 and S3 in the Supplementary File).

As was expected from results discussed above, the HC₅ values predicted from SSD means and SDs that were estimated on the basis of models with the 3-species mean/SD and the best models developed with the 3-species mean/SD plus descriptors were highly correlated with the HC₅ values estimated on the basis of the measured toxicity data (r² ≥ 0.95; Fig. 3), unlike those based on the best models with descriptors alone (r² = 0.67). Also, in the HC₅ predictions based on those models with the 3-species mean/SD, only one chemical (chloroform) was outside a factor of 10 compared with the HC₅ values derived from the measured data, whereas 15 chemicals were outside the range in the prediction based on models with descriptors alone (Fig. 3).

Discussion

Our results based on SSDs derived for 60 chemicals clearly showed that the models developed on the basis of descriptors alone had limited ability to predict the mean and SD of SSD and thereby HC₅; for example, this is particularly apparent from our finding that, when descriptors alone were used, the predicted HC₅ values for 25% of chemicals deviated by more than a factor of 10 from HC₅ values estimated on the basis of the measured toxicity data (Fig. 3). These results are generally consistent with the conclusion of Hoondert et al. (2019) that uncertainties associated with prediction of the mean and SD of SSDs and HC₅ by using descriptors may be high (see Fig. S3 of Hoondert et al., 2019 for the HC₅ prediction). However, compared with our results, the variations not explained by the models were somewhat larger in the work of Hoondert et al. (2019). This difference was obvious from the r² values of their best models (0.56 and 0.29, respectively; see also Iwasaki & Hayashi, 2020), even though the predictors that they used were similar to ours, with the exception of the two binary variables (Class 3 and Class 4). It is difficult to determine the underlying reasons for this difference in predictive ability. Two apparent differences were the quantity of data and whether the quality of each of the experimentally obtained toxicity values was examined. Hoondert et al. (2019) analyzed SSDs based on acute toxicity data for 558 chemicals, which was approximately 9 times the number of chemicals considered in the present study; however, they did not take the quality of each of the experimentally obtained toxicity values into consideration although they differentiated the quality of SSDs based on the inclusion of extrapolated toxicity data. Despite the clear trade-off, both kinds of data should be valuable in developing predictive models that are well tested. There is a need, however, for further testing of the utility of 3-species predictors with large datasets.

Compared with the models with descriptors alone, the addition of a 3-species mean/SD markedly improved the predictions of the mean and SD of SSD (see Table 3), as well as of HC₅. These results suggest that using not only descriptors but also the mean and SD of toxicity data for three species, each of which belongs to one of the three commonly tested biological groups (algae, crustaceans, and fish), is a promising way of producing a more reliable SSD derivation in cases where limited ecotoxicity data are available. Douziech et al. (2020) also demonstrated that HC₅ values estimated on the basis of acute toxicity data for three species of algae, daphnia, and fish were mostly within a factor of 10 of HC₅ values based on all measured toxicity values. Although the uncertainties associated with our model estimates of HC₅ were relatively small—particularly for those obtained by including the 3-species mean/SD as a predictor (see the x-axis error bars in Fig. 3), this prediction assumed that the means and SDs of SSDs calculated from the measured toxicity data for at least 8 species were accurate. Caution is therefore required in the use of such model-based estimates in ecological risk assessments because of the inherent uncertainties in HC₅ values based on measured toxicity data (y-axis error bars in Fig. 3; see also de Zwart 2002 and Hiki & Iwasaki, 2020) as well as the large uncertainties in HC₅ values estimated on the basis of 3 species toxicity data (Douziech et al., 2020).

The binary predictor, Class 4 (specifically acting chemicals), was included in all four best models, with relatively large negative full-model-averaged standardized parameter estimates in the models predicting SSD means and relatively large positive full-model-averaged standardized parameter estimates in the models predicting SSD SDs (see Table 2). These results support previous empirical evidence that specifically acting chemicals are generally more toxic than other classes of chemicals, and that inter-species sensitivity to these chemicals varies more widely than that to other classes of chemicals (Vaal et al., 1997; Hendriks et al., 2013; Hiki & Iwasaki, 2020); they also imply that such patterns cannot be fully captured by using only the 3-species mean and SD. The variations in toxicity values within the biological groups appeared to be relatively high for Class 4 chemicals (Fig. S1), and this may have led to the importance of Class 4 in the models that we developed. However, it is difficult to test this hypothesis robustly owing to the limited number of Class 4 chemicals in our study (a total of 7 chemicals). Another binary predictor, Class 3 (reactive chemicals), was selected in the best model with descriptors alone for the SSD mean, and not in the other best models, suggesting that inclusion of the 3-species mean could play an alternative role for Class 3. The predictive abilities of the best models developed with the 3-species mean/SD plus descriptors were similar among Classes 1–5 based on the Verhaar mode-of-action classification (see Fig. S2 for the differences between the predicted and measured means and SDs of the SSDs).

Log K_OW (or log P) is commonly used in QSAR (quantitative structure–activity relationship) models to predict ecotoxicity (Furuhama et al., 2010; Mayo-Bean et al., 2017), and it was indeed included in the best model with descriptors alone for SSD mean. Log VP was included in the best model with the 3-species SD plus descriptors (Table 2), and its regression coefficient was positive but non-significant (P = 0.072). This may indicate that the increased variation in species sensitivity (i.e., the SD of SSD) was associated with the difficulty in toxicity testing of the more volatile chemicals. However, a clear positive relationship between the SD of SSD and log VP was not apparent owing to the large variation in the SD (Table S1). Note that although use of estimated values of these descriptors is efficient, particularly if the number of chemicals of concern is large, there are uncertainties associated with the estimated values (US EPA, 2007; Enoch et al., 2008) that may have affected the development of the model and the accuracy of its predictions. Further tests of the developed models with experimentally obtained values of descriptors would therefore be valuable, even though experimental results are associated with uncertainties that can vary as a function of experimental conditions.

Conclusions

We demonstrated here that the use of measured toxicity data for three species selected from the three commonly tested biological groups (algae, crustaceans, and fish) markedly improved the model predictions for the means and SDs of log-normal SSDs based on acute toxicity. Furthermore, the HC₅s in SSDs predicted by using such models were generally within a factor of 10 of the values based only on measured toxicity data for at least 8 species. Therefore, the use of 3-species toxicity data could be promising for predicting the HC₅ values from which PNECs are frequently derived in ecological risk assessments. Note that our study was limited to log-normal SSDs based on acute toxicity data for 60 chemicals, whereas PNECs for long-term exposure to chemicals are derived from SSDs based on chronic toxicity data. However, because, for a large proportion of chemicals, collecting chronic toxicity data for three species would be much more difficult and less feasible than collecting acute toxicity data for three species (Hoondert et al., 2019; Posthuma et al., 2019), other prediction approaches, including the use of acute to chronic ratios for SSDs (De Zwart, 2002; Hiki & Iwasaki, 2020) and in silico predictive models (Dyer et al., 2008; Douziech et al., 2020; Takata et al., 2020), could be worth pursuing.

Supplemental Information