State of the Science on Metal Bioavailability Modeling: Introduction to the Outcome of a Society of Environmental Toxicology and Chemistry Technical Workshop

Abstract

A Society of Environmental Toxicology and Chemistry technical workshop was held in December 2017 to critically evaluate the current state of the science of metal bioavailability modeling. The availability of mechanistic models such as the biotic ligand model and the rapid development of empirical models such as multiple linear regressions means that choices are available in terms of bioavailability normalization approaches that can be used in metal risk assessments and the development of risk-based protective values for aquatic life. A key goal of the workshop was to provide potential users of metal bioavailability models with the information required to make appropriate decisions when choosing among mechanistic and empirical models. Workshop participants focused on the state of the science of metal bioavailability modeling, mechanistic and empirical model frameworks, validation of bioavailability models, and application of bioavailability models in risk-based decision-making approaches. The output of this workshop provides the necessary scientific information to incorporate bioavailability normalization in regulations pertaining to metals in freshwater systems.

INTRODUCTION

Accounting for the influence of water chemistry on the ecotoxicity of metals to aquatic life is well recognized in risk-based regulatory approaches (Bergman and Dorward-King 1997; US Environmental Protection Agency 2000). The need for such approaches is demonstrated by the large (>30-fold) differences in intraspecies sensitivity that can be observed when testing typical surface waters that show differences in naturally occurring water chemistry parameters. Initial studies, dating to the 1930s, primarily focused on the influence of hardness cations (Ca²⁺ and Mg²⁺; e.g., Erichsen Jones 1935). Current awareness of the water chemistry parameters affecting metal ecotoxicity goes beyond water hardness and has highlighted the need to account for influences of parameters like dissolved organic carbon (DOC) and pH (Pagenkopf 1983; Allen and Hansen 1996; Di Toro et al. 2001; Merrington et al. 2016).

The mechanisms by which water chemistry constituents influence metal ecotoxicity are well understood, as are methods that can account for this influence. The biotic ligand model (BLM) and related mechanistic-based modeling approaches were developed in the early 2000s (Di Toro et al. 2001) and have been used in the determination of freshwater copper water quality criteria by the US Environmental Protection Agency in 2007 (US Environmental Protection Agency 2007) and in the determination of nickel environmental quality standards by the European Commission in 2013 (European Commission 2010). A key scientific value of the BLM framework is that it performs 2 equally challenging functions (Paquin et al. 2002). First, the underlying speciation module calculates metal speciation for a given water sample as a function of interactions between metals and multiple water chemistry constituents (e.g., pH, DOC, and hardness). Next, the model predicts the ecotoxicological consequences of the interaction of the free metal ion activity to aquatic organisms. For standard species (e.g., Daphnia magna), BLMs have been shown to explain toxicity to other taxonomically unrelated species (e.g., Deleebeeck et al. 2007). These attributes allow for bioavailability correction to be incorporated into environmental risk assessment of metals and specifically the determination of values that are intended to protect aquatic ecosystems.

However, BLMs and related mechanistic models are not simple to use. They require a full range of water chemistry data and experience with dedicated BLM software. These challenges have contributed to a lack of widespread use of mechanistic models and have triggered the exploration of alternative methods (Merrington et al. 2016).

Recently, interest in empirically based models as alternatives to BLM-based approaches has emerged. The most common empirical approach is the use of multiple linear regression (MLR) models, which are essentially statistical relationships between toxicity endpoints (e.g., growth impairment, reproduction) and water chemistry parameters (e.g., hardness, DOC, pH). Therefore, MLR models do not require an independent determination of metal speciation and hence do not require the extensive water chemistry data required to make these calculations. Recent papers (Brix et al. 2017; De Forest et al. 2018) have illustrated how MLRs can be developed in ways that are consistent with the mechanistic understanding developed in BLM determinations and how these empirical models can be used to derive criteria and standards.

Despite the availability of mechanistic (e.g., BLMs) and empirical (e.g., MLRs) modeling approaches for incorporating metal bioavailability into the development of values that are protective of aquatic life, little guidance is available for users to choose among different modeling options. Traditionally, model performance has been evaluated by comparing predicted and observed toxicity, and an informal metric of a 2-fold difference has been used to indicate that a model is sufficiently robust. The 2-fold agreement was first used in a specific context to show the degree to which a BLM decreased the intraspecific variability for 1 species using 1 set of ecotoxicity data. Other decisions enter into the evaluation of a model and how to choose among more than 1 model that meets the factor-of-2 rule of thumb.

Considering the current availability of several scientifically defensible approaches, a Society of Environmental Toxicology and Chemistry (SETAC) technical workshop was held in December 2017 to take stock of the current state of the science of metal bioavailability models, to evaluate the performance of the models, and to identify best practices in the use of these models in the determination and application of bioavailability-based effect concentrations for metals that are intended to protect aquatic life (e.g., criteria and standards). The following 5 papers represent the outcome of discussions that took place prior to, during, and after the SETAC technical workshop.

WORKSHOP APPROACH AND PARTICIPANTS

Forty-one experts from academia (n = 13), business (n = 17), and government sectors (n = 11) in 7 countries from North America, Europe, and Australia participated in the week-long workshop (Figure 1). Postdoctoral researchers and 1 graduate student were among the participants and provided an early career perspective.

FIGURE 1:

Participants at the SETAC Technical Workshop Bioavailability-Based Aquatic Toxicity Models for Metals, held in Pensacola, Florida, USA, in December 2017.

Participants were divided among 5 workgroups that addressed the following issues: 1) Bioavailability assessment of metals in aquatic environments: A historical review (Workgroup chairs: Bill Adams, Red Cap Consulting, and Dave Mount, US Environmental Protection Agency). 2) Metal bioavailability models: Current status, lessons learned, considerations for regulatory use, and the path forward (Workgroup chairs: Chris Wood, University of British Columbia, and Chris Mebane, US Geological Survey). 3) Development of empirical bioavailability models for metals (Workgroup chairs: Kevin Brix, Ecotox, and Russ Erickson, US Environmental Protection Agency). 4) Validation of bioavailability-based freshwater toxicity models for metals (Joe Meyer, Applied Limnology Professionals, and Emily Garman, NiPERA Inc.). 5) Derivation and application of thresholds for metals using bioavailability-based approaches (Jenny Stauber, Commonwealth Scientific and Industrial Research Organization, and Eric Van Genderen, International Zinc Association).

Leading up to the workshop, workgroups conducted a series of conference calls to organize each group’s charge questions and overall goals. Once in Pensacola, the workshop began with a series of “scene-setting” talks on past and current approaches for incorporating bioavailability into criteria development in the United States and elsewhere in the world and illustrations on how mechanistic and empirical models can be used for development of protective values for aquatic life (pvals). Most of the remaining time was spent in workgroup discussions, with a daily plenary “check-in” of progress toward day’s end and discussions of high-level cross-cutting issues.

WORKGROUP FINDINGS AND RECOMMENDATIONS

Workgroup 1: Bioavailability assessment of metals in aquatic environments: A historical review (Adams et al. 2020)

Adams et al. (2020) outline the chronology of the development of bioavailability concepts and the subsequent incorporation of these concepts in regulatory decision-making. As Adams et al. (2020) show, reports that water chemistry affects metal toxicity date to the 1930s. Since that early exploratory period, methods have transitioned from relatively simple hardness-based corrections of metal ecotoxicity to complex modeling approaches that account for the full complement of water chemistry parameters that influence metal toxicity. The most prominent of these modeling approaches was the BLM, with BLMs now being available for most industrial metals and being used for regulatory purposes in the United States and elsewhere. Adams et al. (2020) acknowledge the complexity associated with the BLM as 1 reason for the recent emergence of empirically based approaches like MLRs.

In general, the regulatory application of bioavailability models has lagged behind scientific development. This trend appears to be changing because several global regulatory agencies are currently embracing MLR concepts, even though these approaches have only recently appeared in the scientific literature. From a broader perspective, Adams et al. (2020) conclude that the activity on developing mechanistic and empirical models highlights the need for robust, objective, and comprehensive approaches on model evaluation and selection.

Workgroup 2: Metal bioavailability models: Current status, lessons learned, considerations for regulatory use, and the path forward (Mebane et al. 2020)

Mebane et al. (2020) focus on the modeling approaches that are most familiar, the mechanistic-based models based on the BLM concept, which was first conceived in the late 1990s, when the processes of predicting metal speciation in the dissolved phase and determining the toxicological consequences of metal accumulation at the site of toxic action were unified in a single computational framework. The seminal publications of DiToro et al. (2001) and Paquin et al. (2002) describe the BLM framework, which emphasized methods supporting mechanistic approaches to determine model parameters, such as actual measurement of metal accumulation as a function of water chemistry at the purported “biotic ligand” and site of toxicity, where the gills of freshwater fish were an example (Playle 1998). Later efforts utilized toxicity tests conducted across wide ranges of water chemistry to derive these parameters (e.g., De Schamphelaere and Janssen 2002). Mebane et al. (2020) make a distinction between these approaches from a methodological perspective but conclude that they serve the same purpose of predicting metal toxicity within a mechanistic framework.

Mebane et al. (2020) critically evaluate several issues pertaining to the breadth of mechanistic models. Chief among these is the behavior of models in waters of “extreme chemistry” (e.g., waters with pH, hardness, and DOC that are outside of conditions used to construct the models and outside of conditions in typical natural waters). The importance of water physical and chemical parameters that are typically not considered, including temperature, Fe, Al, and nutrients, is also explored in terms of their influence on model outcome.

Mebane et al. (2020) conclude with a series of principles and recommendations for developing mechanistically based models and for considering how these models can be used to predict metal toxicity. Multivariate test designs (e.g., testing natural waters varying in water chemistry) can help identify important parameters. However, determining the influence of specific water chemistry parameters (e.g., Ca²⁺) with precision is best performed by using a univariate approach (e.g., tests in which one parameter varies and others are kept constant). Mechanistic bioavailability models are based on observations from ecotoxicity tests. To this end, the organisms used in model development should generally not be tested in waters with chemistries outside of their physiological tolerances. Effects of pH on metal toxicity are unique compared with other parameters and often show nonlinear relationships. Models that include nonlinear pH relationships and linear relationships of other parameters (e.g., hardness) are the best approach in these circumstances.

Workgroup 3: Development of empirical bioavailability models for metals (Brix et al. 2020)

Empirical modeling approaches include water hardness correction, which has been incorporated in water standards since the mid-1980s. However, the latest empirical approaches are new and reflect a response to the complexity associated with using and implementing mechanistically based approaches like the BLM. A common empirical approach that has emerged recently is the use is of MLRs. Brix et al. (2020) describe the approach for developing an MLR, which begins with a species-specific data set developed over a range of water chemistries. The outcome is a regression equation in which water chemistry parameters are independent variables that predict the toxicity of the test species in question. One of the more attractive features of the MLR approach is that the influence of water chemistry on metal toxicity is directly measured, as opposed to being based on a calculated metal ion activity, which is the case of mechanistic models like the BLM.

In the literature, MLRs have appeared with increasing frequency, and excellent examples of their development and proposed use are available for Cu (Brix et al. 2017) and Al (De Forest et al. 2018). The importance of these developments is demonstrated by the recent use of an MLR to serve as the basis for determining the final aquatic life ambient water quality criteria for Al (US Environmental Protection Agency 2018).

Despite these advancements, little guidance on best practices for developing MLRs is available. Brix et al. (2020) address this need by providing recommendations on the identification of water chemistry parameters to consider, the appropriate range of water chemistry parameters that should be addressed for the model to fulfill its purpose, and critical considerations in terms of evaluating the statistical robustness of resulting models. In addition, modeling attributes that are specific to MLRs are explained, including the importance of identifying interactions between variables, goodness-of-fit tests appropriate for regression modeling, and the need to distinguish between linear and nonlinear relationships. These are novel contributions that will assist in the development and evaluation of empirical models.

Workgroup 4: Validation of bioavailability-based toxicity models for metals (Garman et al. 2020)

Models used to predict metal toxicity as a function of water chemistry should undergo validation before they can be used for regulatory purposes. However, there is little guidance available to validate existing mechanistic models like BLMs or the emerging empirically based approaches like MLRs. Garman et al. (2020) outline what validation means in the context of metal bioavailability models and what it does not mean. For the purposes of the present exercise, “validation” refers to the ability of a species-specific model to accurately predict an ecotoxicity endpoint using the model’s identified water chemistry parameters as independent variables. “Validation” does not refer to the level of ecological protection offered by a bioavailability-based pval because the protectiveness of a pval is influenced by factors beyond the precision of metal toxicity predictions (e.g., the ecotoxicity data that are used, the degree of protection that is chosen, statistical methods used to determine the pval).

Essentially, the validation approach described by Garman et al. (2020) assesses the appropriateness, relevance, and accuracy of metal bioavailability models. Appropriateness and relevance are judged based on qualitative criteria, whereas accuracy is the focus of newly developed, quantitative evaluation approaches. To date, the rule of thumb in terms of evaluating model accuracy has been to ask if the model predictions are within a factor of 2 of observed toxicity measurements. However, the factor-of-2 rule can overlook important biases that can be revealed by employing different statistics. In particular, the workgroup recommends the use of residuals as measurements of statistical characteristics (e.g., slope, mean) to examine bias in the model performance. This analysis can reveal over- or underpredictions of toxicity, systematic trends in bias (e.g., showing that over- or underprotection is related to the magnitude of the measured endpoint), and relationships between bias and specific water chemistry parameters. Outcomes of the analyses proposed by Garman et al. (2020) can be used by model developers to refine models and by model users when choosing among more than one potentially suitable model.

Workgroup 5: Derivation and application of thresholds for metals using bioavailability-based approaches (Van Genderen et al. 2020)

“Application” can mean different things to different audiences, and it was important to define its meaning within the context of the workshop. Van Genderen et al. (2020) described 2 areas where bioavailability models can be applied in a regulatory context. The first area is the application of models for the derivation of pvals such as aquatic life standards/criteria. The second area describes how bioavailability models can be used in conjunction with pvals to address specific risk-based goals, such as the setting of permit limits. Recommendations are provided on screening among the multitude of available ecotoxicity data to select those that can be used to derive bioavailability-based pvals. Not all studies in the literature report the water chemistry data that are required by bioavailability models. The workgroup also provided recommendations for selecting the most appropriate model for pval derivation. This is a particularly important advancement given the availability of multiple models for a given metal that may differ in terms of precision and ease of use. Van Genderen et al. (2020) identify the following characteristics of a model that should be used in evaluations: 1) The representativeness of the water chemistry parameters used in model development to the range of water chemistry in the region where the model is to be applied and the range of taxonomic groups for which the bioavailability data of the model applies. 2) The level of input, which refers to the breadth of water chemistry parameters that are required to operate the model. 3) The accuracy of the model as demonstrated by model calibration and validation results, where the implication is that any model that is to be considered for regulatory use should have these data. 4) Ease of use, which includes the level of training required to use the model, compatibility of the model with commonly available computer software, and similarity between the output of the model and what is required by the regulatory framework in question.

A case study is provided in which the bioavailability models that are available for zinc are compared. The models included simple hardness corrections, MLRs, and a BLM framework. Most comparisons are based on qualitative scoring, such as whether or not certain taxonomic groups are covered in the modeling data. In terms of quantitative comparisons, the workgroup recommended the use of residual factors (RF), which represents the fraction or percentage of predictions that fall below a given target. For example, an RF₂ would be used to identify the proportion of predictions that fall within a 2-fold prediction. In this specific example, the RF₂ ranged from 0.19 for hardness correction (i.e., 19% of predictions are within a factor of 2 of the observed toxicity value) to 0.87 for BLMs (i.e., 88% within a factor of 2). The outcome of this comparison provides practical and objective information that allows the user to decide which model is most appropriate for a given situation.

CONCLUSIONS AND NEXT STEPS

The SETAC technical workshop resulted in high-level recommendations regarding the development of metal bioavailability models and their use. First, all bioavailability models should be informed by a mechanistic understanding of metal toxicity and of metal speciation. Second, the development of simplified tools is feasible, as long as the tools have mechanistic links. Third, all models should undergo qualitative and quantitative validation and be applied within appropriate application ranges of water chemistry. Finally, different models can be used for different situations; the critical aspect is that the choice of the most appropriate model needs to be transparently communicated.

In conclusion, the outcomes of the workshop that are communicated in the following series of papers provide a solid basis for developing and applying bioavailability models for metals that reflect the state of the science, are flexible in terms of how they reflect influences of site-specific water chemistry, and are straightforward in terms of their application. Given that the scientific principles governing metal bioavailability are universal, the conclusions and recommendations from the workshop can support expanded incorporation of metal bioavailability information into regulatory frameworks around the world.