Environmental Toxicology and Chemistry, Vol. 34, No. 4, pp. 821–832, 2015 # 2014 SETAC Printed in the USA

Metal Mixture Modeling Evaluation METAL MIXTURE MODELING EVALUATION PROJECT: 3. LESSONS LEARNED AND STEPS FORWARD KEVIN J. FARLEY*y and JOSEPH S. MEYERz yDepartment of Civil and Environmental Engineering, Manhattan College, Riverdale, New York, USA zARCADIS U.S., Lakewood, Colorado, USA (Submitted 29 April 2014; Returned for Revision 19 August 2014; Accepted 1 December 2014) Abstract: A comparison of 4 metal mixture toxicity models (that were based on the biotic ligand model [BLM] and the Windermere humic aqueous model using the toxicity function [WHAM-FTOX]) was presented in a previous paper. In the present study, a streamlined version of the 4 models was developed and applied to multiple data sets and test conditions to examine key assumptions and calibration strategies that are crucial in modeling metal mixture toxicity. Results show that 1) a single binding site on or in the organism was a useful and oftentimes sufficient framework for predicting metal toxicity; 2) a linear free energy relationship (LFER) for bidentate binding of metals and cations to the biotic ligand provided a good first estimate of binding coefficients; 3) although adjustments in metal binding coefficients or adjustments in chemical potency factors can both be used in model calibration for single-metal exposures, changing metal binding coefficients or chemical potency factors had different effects on model predictions for metal mixtures; and 4) selection of a mixture toxicity model (based on concentration addition or independent action) was important in predicting metal mixture toxicity. Moving forward, efforts should focus on reducing uncertainties in model calibration, including development of better methods to characterize metal binding to toxicologically active binding sites, conducting targeted exposure studies to advance the understanding of metal mixture toxicity, and further developing LFERs and other tools to help constrain the model calibration. Environ Toxicol Chem 2015;34:821–832. # 2014 SETAC Keywords: Biotic ligand model WHAM-FTOX

Concentration addition

Metal bioavailability

Metal toxicity

Independent action

binding to “biotic ligand[s]” [in the AIST, USGS, and HDR models], or by using the WHAM-FTOX approach, with metal binding to humic acid serving as a proxy for nonspecific metal accumulation by the organism [in the CEH model]); and the specification of metal potency factors, log-logit response functions, or a linear-threshold response function to relate metal accumulation to biological response. Major differences in the models were attributed largely to various modeling assumptions (e.g., single vs multiple types of binding sites on or in the organism) and specific calibration strategies, which affected the selection of binding coefficients for major cations and metals, metal-potency factors, and toxicity-response parameters. Comparison of the AIST, USGS, CEH, and HDR models provided unique insights into model development and helped in furthering the understanding of metal mixture toxicity. The strengths and potential weaknesses of assumptions and calibration strategies of each modeling approach, however, were sometimes difficult to evaluate because of the complexities of the models, the interrelationship of many of the model parameters, and the flexibility in fitting models to data sets with only metal-exposure concentrations, water chemistry measurements, and observed organism responses. The purpose of the present study therefore is to examine the key assumptions and calibration strategies that are most crucial in modeling the available metal mixture toxicity data sets using a streamlined version of the 4 modeling approaches. A description of the streamlined model is presented and is followed by the application of the model to various metal mixture toxicity data sets in a series of model-calibration and model-validation exercises. As part of this effort, specific model assumptions and calibration procedures are examined, model results are interpreted, and lessons learned from the larger

INTRODUCTION

A Metal Mixture Modeling Evaluation (MMME) project was initiated to evaluate the current capabilities of 3 biotic ligand models (BLMs) and the Windermere humic aqueous model using the toxicity function (WHAM-FTOX) in predicting metal mixture toxicity and to promote the continued development of various modeling approaches [1]. Under this initiative, models were developed/refined and tested by researchers from the National Institute of Advanced Industrial Science and Technology (AIST) in Ibaraki, Japan [2]; the US Geological Survey (USGS) in Seattle, Washington, and Boise, Idaho, USA [3–5]; HDR j HydroQual, Inc. (HDR) in East Syracuse, New York, USA [6]; and the Centre for Ecology and Hydrology (CEH) of the Natural Environment Research Council in Lancaster, United Kingdom [7]. Summaries and a comparison of the 4 models, as they were presented at an MMME workshop in Brussels, Belgium, in May 2012 were provided in a previous study [3]. Based on that evaluation, the 4 models were found to be similar in their overall structure (Figure 1). The 4 models consisted of the following major components: a chemical speciation calculation to compute the free-ion activities of metals (and major cations) based on competitive binding to inorganic ligands and dissolved organic matter (DOM) using various versions of WHAM; an evaluation of competitive binding of metals and major cations to 1 or more binding sites on or in the organism using conventional competitive equilibrium chemistry (this was described either by considering metal All Supplemental Data may be found in the online version of this article. * Address correspondence to [email protected] Published online 5 December 2014 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/etc.2837 821

822

Environ Toxicol Chem 34, 2015

Figure 1. General structure of the 3 biotic ligand models (BLMs) and Windermere humic aqueous model using the toxicity function (WHAMFTOX) model used in predicting metal mixture toxicity. Metal accumulation on the biotic ligand or the humic acid proxy are related to toxicity using potency factors or toxicity response functions. DOM ¼ dissolved organic matter; BL ¼ biotic ligand site(s); HA ¼ humic acid, which is used as a proxy in computing nonspecific metal accumulation by the organism in WHAM-FTOX.

modeling efforts are highlighted. Specific questions that are addressed include the following: 1) Are single-site models sufficient for predicting metal mixture toxicity? 2) Can a linear free energy relationship (LFER) be used to constrain the selection of binding coefficients for sites on or in the organism? 3) Can toxic responses be correlated to bound metal concentrations without the need to specify metal-specific potency factors (or metal-specific response intercepts or slopes)? 4) Is it more appropriate to assume “joint similar action” (also referred to as “concentration addition”) or “joint independent action” (also referred to as “response addition”) in predicting metal mixture toxicity? MODEL DESCRIPTION

The general structure of the streamlined model that is used in the present study most closely follows the modeling approach presented by the USGS [3]. This selection is not meant to imply that the USGS model is superior to the other modeling approaches. Rather, the USGS model appears to provide a more flexible framework for evaluating the key assumptions and calibration strategies in the 4 models because it is based on the BLM approach but includes several key elements from WHAMFTOX. Details of the streamlined model are given below. WHAM calculations

In the 4 modeling studies, free ion activities of metals and major ions were computed using WHAM V (HDR), WHAM VI (CEH), or WHAM VII (AIST, USGS). For our evaluations, WHAM VII [8,9] was selected for the streamlined model because it is the most updated version of WHAM and has been tested extensively for proton, cation, and metal binding to DOM. In addition, WHAM VII has a strong chemical basis and in part employs a chemical similarity approach in assigning binding coefficients (log KM values) for major cations and metals [10].

MMME Lessons Learned

The WHAM VII calculations for each metal toxicity test were performed as follows: Dissolved concentrations of major ions (Naþ, Kþ, Ca2þ, Mg2þ, Cl–, SO42–) and metals (Cd, Cu, Ni, Pb, Zn) were inputted into the model along with the dissolved organic carbon (DOC) concentration, temperature, pH, and alkalinity (with pH and alkalinity used to compute bicarbonate and carbonate concentrations). Following the USGS approach [3], DOM was assumed to equal 2 times the reported DOC concentration, and DOM was assumed to be 65% active with 10% of the active fraction defined by WHAM humic acid and 90% defined by WHAM fulvic acid. Thus, conversions from DOC concentration ([DOC], in mg L1) to WHAM inputs for humic acid and fulvic acid (in g L1) were: humic acid ¼ 2  0.65  0.1  0.001  [DOC] and fulvic acid ¼ 2  0.65  0.9  0.001  [DOC]. For toxicity tests conducted in field-collected water samples, the free-ion activities of Al and Fe(III) were computed internally in the model based on assumed equilibrium with aluminum and iron hydroxides using solubility relationships [11,12]. For tests conducted in laboratory water, Al and Fe(III) concentrations were considered negligible. Metal binding

A variety of approaches were used in the previous models to describe binding of major cations and metals to binding sites on or in the organism [3]. These included a single biotic ligand site with adjustable cation/metal binding coefficients (in the AIST and USGS models), a multi-site biotic ligand where all metals can bind to each site but only 1 metal is toxic at its unique toxicity site (in the HDR model), and a distribution of monodentate, bidentate, and tridentate binding sites to describe nonspecific metal accumulation to a humic acid proxy for nonspecific metal accumulation by the organism (in the CEH model). Although there are sound justifications for selecting the multi-site approaches, a single biotic ligand site was selected for the streamlined model to help constrain the selection of modelcalibration coefficients and to simplify the interpretation of model results. To further constrain model calibration, an LFER was used in setting log KM values for major cation and free metal (M2þ) binding to the biotic ligand site, where KM is the equilibrium binding constant. Selection of an appropriate LFER, however, can be somewhat complicated because binding sites on or in an organism are likely to be composed of various functional groups (including combinations of carboxyl, phenol, amine, and sulfur groups) that have different affinities for metals. Although binding to soft ligands (e.g., amine and sulfur groups) is more likely to be important on the biotic ligand, LFERs for soft ligands were not available at the time of the present study. Therefore, the Irving-Rossotti LFER for bidentate binding to negatively charged oxygen donor groups [13] was selected to describe cation/metal binding to the biotic ligand in the initial calibration of the model. This approach allows all binding coefficients (log KM values) for major cations and metals to be calculated directly from the acidity constants for the 2 donor groups (pKa1, pKa2). In applying the Irving-Rossotti LFER, the first acidity constant (pKa1) was set equal to 3.6 based on the proton dissociation constant reported for carboxylic functional groups in the cell walls of algae [14]. The second acidity constant (pKa2) was then the only adjustable parameter for calculating binding of metals, major cations, and protons on the biotic ligand. In addition to the binding of free metal ions (M2þ) to the biotic ligand, binding of the hydrolyzed metal (MOHþ) is also considered in some BLMs for individual metals [15] and for metal mixtures [3,6]. Following the approach in the WHAM

K.J. Farley and J.S. Meyer

Environ Toxicol Chem 34, 2015

models [8], log KM values for binding of hydrolyzed metals to the biotic ligand were set equal to their respective log KM values for the free-metal ions. Toxicity response

A variety of approaches were also used in the previous models to relate metal accumulation on the biotic ligand to reductions in algal growth, Daphnia magna mortality, and rainbow trout (Oncorhynchus mykiss) and cutthroat trout (O. clarkii) mortality. These approaches included several different logit and linear-threshold functions to correlate the accumulation (or potency) of individual metals on the binding sites to toxic response, combined with metal mixture models based on concentration addition and independent action (see Farley et al. [3] for details). Three permutations of these approaches were considered in the streamlined model. The first approach, concentration addition, is based on the assumption that all metals elicit an equivalent toxic effect (e.g., Ca uptake inhibition and hypocalcemia) and that mixture toxicity is related to the total accumulated metal on the biotic ligand [16]. For the concentration addition calculation, metal coverage on the biotic ligand is related to growth reduction (or mortality) using a 2-parameter log-logit function R¼

1 1 þ eða þ b ln uM Þ

ð1Þ

where R is the proportional response (growth reduction or mortality), a and b are the logit intercept and slope of the toxicity-response curve, and uM is the fractional coverage of metal on the biotic ligand. For metal mixtures, uM is based on total metal bound to the biotic ligand. The second approach, TOX addition, is a special version of concentration addition that is used when each metal is assumed to exhibit a different potency on the biotic ligand. For the TOX addition approach, the toxicity of a single metal is related to the metal potency factor (ai) times the fractional coverage of metal on the biotic ligand (ui) as TOX i ¼ ai ui

ð2Þ

Growth reduction (or mortality) is then calculated using the 2parameter log-logit function (Equation 1) by replacing uM with TOXi. For metal mixtures, the toxic potency of each metal is assumed to be additive and is summed as TOX ¼ S TOX i ¼ S ai ui i

i

ð3Þ

and the growth reduction or fractional mortality is again determined using Equation 1 by replacing uM with TOX.

823

Finally, the joint independent action (or response addition) approach considers that each metal elicits an effect at a different binding site and that mixture toxicity can be determined from the product of the expected individual-metal growth or survival proportions. In this case, the fractional coverage of each metal on its toxicologically active biotic ligand is again related to the growth reduction (or mortality) using the 2-parameter logit function (Equation 1). For metal mixtures, growth reduction (or mortality) is determined from the expected proportional growth reduction or mortality that would be caused independently by each metal (Ri) as R ¼ 1  P ð1  Ri Þ i

ð4Þ

In our application of independent action, all metals are allowed to bind to each site, and metal binding to all sites is described by a single set of log KM values. METHODS

Model calibration: Single metal exposures

The streamlined model was calibrated using single-metal exposure data from 3 of the larger MMME calibration data sets (Table 1). Zinc-only exposure data for growth reduction of Pseudokirchneriella subcapitata (a freshwater green algae, Index 8) was first considered. In this procedure, the chemical potency factor for Zn (aZn) was set equal to 1.0, and the remaining model coefficients (pKa2, a, b) were obtained by fitting the Zn-only exposure data at pH 6.0 and pH 8.4. This was accomplished using an iterative approach, in which log KM values for Zn and the competing cations (Mg2þ, Ca2þ) were calculated at each step in the iteration by using the selected pKa2 value and the Irving-Rossotti LFER (Supplemental Data, Equation S1 and Table S1). Final calibration was based on graphical comparisons of calculated and observed reduction in population growth. Zinc-only exposure data for the remaining 2 calibration data sets (Index 4, Index 6) were then considered. To help constrain model calibration for the Index 4 and Index 6 data sets, the pKa2 value (and the corresponding LFER-calculated log KM values for Zn, Mg, and Ca) was taken directly from the initial model calibration for growth reduction of P. subcapitata. The aZn value was again set equal to 1.0, and the remaining calibration parameters (a, b) for Index 4 and Index 6 were obtained by fitting the Zn-only exposure data using the iterative approach described. Model calibration to single-metal exposure data for other metals was considered first for the P. subcapitata data set (Index 8). As a starting point, pKa2, a, and b values from the Zn calibration were applied to the other single-metal exposure data

Table 1. Metal Mixture Modeling Evaluation project data sets used in model calibration and validationa Index 8 4 6 V-1 V-3 a

Species

Metal mixture/water type

Endpoint

No. of exposuresb

Source

Pseudokirchneriella subcapitata Daphnia magna Rainbow trout/cutthroat troutc D. magna Rainbow trout

Cd-Cu-Ni-Zn/ Field Cd-Cu, Cu-Zn/ Laboratory Cd-Pb-Zn/ Field Cd-Zn/Laboratory Cd-Cu-Zn/ Laboratory

72-h growth 48-h survival 96-h survival 48-h survival 96-h survival

102/12 387/174 298/71 132/177 72/24

[3] [28] [29] [28] [30]

See Van Genderen et al. [1] for detailed descriptions. Single-metal or reference exposures/mixture exposures. c Rainbow trout (Oncorhynchus mykiss) and cutthroat trout (Oncorhynchus clarkii). b

824

Environ Toxicol Chem 34, 2015

to test whether P. subcapitata growth reduction could be adequately described by the LFER-calculated log KM values and aM values of 1.0. If the model did not provide a reasonable fit to the other single-metal exposure data under the described constraints, model coefficients were further adjusted using 1 of 2 different calibration strategies. In the first calibration strategy, the model was calibrated by adjusting the log KM value (while maintaining all aM values equal to 1.0). In the second calibration strategy, the model was calibrated by adjusting the aM of the metal to a value other than 1.0 (while retaining all of the LFERcalculated log KM values). A similar approach was followed in evaluating other single-metal exposure data for the D. magna (Index 4) and rainbow/cutthroat trout (Index 6) toxicity data. In all cases, final calibration was based on graphical comparisons of calculated and observed growth reduction or mortality. Pearson correlation coefficients (r) also were computed and used as an overall measure of the model calibration. Model predictions: Metal mixtures

Final model coefficients for the single-metal exposures were used directly in predicting mixture toxicity to P. subcapitata, D. magna, and rainbow/cutthroat trout. Three mixture models were considered. These included concentration addition (for which mixture toxicity was assumed to be related to the total accumulated metal on the biotic ligand, using Equation 1), TOX addition (for which each metal was assumed to exhibit a different potency on the biotic ligand, and metal toxicity was considered additive using the TOX function in Equation 3), and independent action (for which mixture toxicity was expressed in terms of a multiplicative function of the responses to the

MMME Lessons Learned

individual metals using Equation 4). Final model coefficients also were tested using 2 independent validation data sets for mortality of D. magna and rainbow trout (see Index V-1 and V-3 in Table 1). RESULTS AND DISCUSSION

Model calibration: Single-metal exposures

Initial calibration of the streamlined model was performed using Zn-only exposure data (with aZn set equal to 1.0). For the P. subcapitata data set, the 3 calibration parameters (pKa2 ¼ 7.5; a ¼ 3.5; b ¼ 1.0) were determined by fitting model response curves to observed growth reductions at pH 6.0 and 8.4 (Figure 2A). The log KM values that were calculated from pKa1, the calibrated pKa2 value, and the Irving-Rossotti LFER (Supplemental Data, Equation S1) are listed in Table 2. The LFER-calculated log KM values from the P. subcapitata calibration were subsequently used in calibrating the model to the Zn-only exposure data for D. magna mortality (Index 4). The aZn was again set equal to 1.0, and the remaining 2 calibration coefficients (a ¼ 27, b ¼ 9) were obtained by fitting the model response curve to observed mortality. A similar calibration was performed using the Zn-only exposure data for the rainbow/cutthroat trout (Index 6). Because all hatchery trout (including rainbow trout from Mt. Lassen Trout Farm in Red Bluff, California, and cutthroat trout and Kootenai strain rainbow trout from Sandpoint Hatchery, Idaho) appeared to exhibit similar sensitivity to metals, they were modeled collectively (a ¼ 16, b ¼ 5). A separate calibration of the model was performed for field-collected cutthroat trout because their

Figure 2. Comparison of model-calculated response curves and observed growth reduction of Pseudokirchneriella subcapitata in field-collected water with spiked metal concentrations (Index 8) based on the adjusted binding coefficient (log KM) calibration listed in Table 3 (with all metal potency factor [aM] values equal to 1.0). Results are shown for single-metal exposures of (A) Zn, (B) Ni, (C) Cu, and (D) Cd. See Table 1 for further description of the data.

K.J. Farley and J.S. Meyer

Environ Toxicol Chem 34, 2015

825

Table 2. Initial model calibration based on Zn-only exposure dataa,b Index 8 4 6 6

Species

No. of Zn-only exposures

a

b

LA50 (Zn)

Pseudokirchneriella subcapitata Daphnia magna Hatchery rainbow/cutthroat troutc Field-collected cutthroat trout

24 42 139 13

3.5 27 16 12

1.0 9 5 5

0.030 0.050 0.041 0.091

Initial calibration is based on first acidity constant (pKa1) and second acidity constant (pKa2) for the bidentate biotic ligand site given as 3.6 and 7.5, respectively; and a chemical potency factor for Zn (aZn) set equal to 1.0. Binding constants (log KM) for cations and metals on the bidentate biotic ligand site as calculated from the Irving-Rossotti linear free energy relationship [13]: log KMg ¼ 4.26, log KCa ¼ 3.23, log KZn ¼ 5.13, log KNi ¼ 5.05, log KPb ¼ 5.74, log KCu ¼ 7.60, log KCd ¼ 4.42. c Rainbow trout (Oncorhynchus mykiss) and cutthroat trout (O. clarkii). a ¼ intercept of the log-logit response function; b ¼ slope of the log-logit response function; LA50 (Zn) ¼ fractional coverage of Zn on the biotic ligand that would result in a 50% response. a

b

toxic responses indicated lower sensitivity to metals. This was achieved by decreasing the log-logit intercept (a ¼ 12) while all other model coefficients were held constant. The calibration parameters for the Zn-only exposure data are listed in Table 2. The fractional coverage of Zn on the biotic ligand that would result in a 50% response ranged from 0.030 for P. subcapitata to 0.091 for field-collected rainbow trout (Table 2). The LFER-calculated log KM values from the initial P. subcapitata calibration were then used along with the toxicityresponse parameters (a, b) from the Zn-only exposure data (Table 2) to test whether the model could adequately describe the observed toxicity for single-metal toxicity tests with other metals. As shown in Figure 2B, the model was able to describe growth reductions for P. subcapitata reasonably well for Nionly exposures at both pH 6.0 and pH 8.4 without any additional adjustment in model coefficients. Application of the model to Cu- and Cd-only exposure tests from the P. subcapitata data set was less successful. Additional adjustment of the initial calibration parameters was therefore performed using 2 calibration strategies: adjusting the log KM values while maintaining all aM values equal to 1.0, or adjusting the aM values while maintaining the LFER-calculated log KM values. Final model fits with adjusted log KM values are shown for Cu and Cd in Figures 2C and 2D. Almost identical model fits were also obtained with adjusted aM values (Supplemental Data, Figure S2). Adjustments to the initial calibration parameters also were required to fit the D. magna mortality data (Index 4) for other single-metal exposures. This calibration was particularly difficult for Cd-only exposures because similar observations in D. magna mortality often corresponded to a wide range of computed accumulations of Cd on the biotic ligand. Because oversaturation of CdCO3(s) in moderately high pH exposure waters may have contributed to some of this variability, test solutions that were suspected of being oversaturated with CdCO3(s) were identified and were not included in the calibration (see Supplemental Data for further details). Final calibration results with the adjusted log KM values are presented for D. magna in Figure 3A. Because the D. magna were exposed to a range of water chemistries (with different pH, hardness, and alkalinity values), calibration results were plotted in terms of mortality versus the calculated fractional coverage of metal on the biotic ligand. Similar results for hatchery rainbow/cutthroat trout are presented in Figure 3B. Finally, almost identical calibration of the D. magna mortality (Index 4) and rainbow/ cutthroat trout mortality (Index 6) was obtained with adjusted aM values (Supplemental Data, Figure S3). Because the various metals exhibit different potencies in this approach, calibration

results with adjusted aM values were plotted in terms of mortality versus the TOXi parameter calculated using Equation 2. Final results from the graphical comparisons discussed and from Pearson correlation coefficients (Table 3) showed that either adjustment of log KM values or adjustment of aM values (or some combination of the 2 approaches) can be used in model calibration for single-metal exposures for data sets that contain only metal-exposure concentrations, water chemistry measurements, and observed organism responses. Therefore, both calibration strategies were considered in applying the model to metal mixtures in the Modeling predictions: Metal mixtures section. Before moving on to the metal mixture toxicity results, however, several important issues should be noted for the calibration of single-metal exposures. First, the ability of the model to describe the effects of single-metal exposures on growth reduction of P. subcapitata over a wide range in pH (Figure 2) was largely attributed to the selection of a bidentate site for metal binding to the biotic ligand. Second, the model’s ability to accurately describe a reversal in pH behavior for Ni toxicity (as shown in Figure 2B, with smaller reductions in algal growth associated with the higher pH) was attributable to the formation of Ni–carbonate complexes and the associated decrease in Ni bioavailability in the higher-pH exposures. Third, oversaturation of CdCO3(s), PbCO3(s), and other solid phases were identified as a concern in toxicity studies using moderately high pH exposures. Fourth, data points for rainbow trout mortality that fell outside of a factor of 2 of the modelpredicted response curve (Figure 3B) were associated with toxicity tests that were conducted at lower pH (0.10) fractional coverage of Cu (the competing metal) on the biotic ligand. Comparisons of model predictions and observed mortality for the hatchery rainbow/cutthroat trout data (Index 6) are presented in Figure 7 following the plotting formats previously described for D. magna (Figure 5). Results for the 4 cases suggest that the effects of Cd-Pb-Zn ternary mixtures on trout mortality may be better described by Case 2 (adjusted log KM values/independent action) and Case 4 (adjusted aM values/ independent action). This finding suggests that joint independent action may provide a better description of rainbow/ cutthroat trout toxicity in Cd-Pb-Zn mixtures. Similar comparisons of model predictions and observed mortality for fieldcollected cutthroat trout data (Index 6) are presented in

Figure S4 in the Supplemental Data. Results were consistent with previous findings for mortality of hatchery rainbow/ cutthroat trout. The calibrated model for D. magna was also applied to a validation study for D. magna mortality in Cd-Zn binary mixtures (Index V-1). Comparison of model predictions and observed mortality for the 4 cases described were similar, with most of the observations falling within a factor of 2 of the model-predicted response curve on the uM, uM(equiv), TOX, or TOX(equiv) scale (Supplemental Data, Figure S5). A more detailed examination of results for D. magna mortality for a test series with a constant Cd concentration of 19.4 mg/L and increasing Zn exposure concentrations showed that independent action was likely to be a better predictor of toxicity for Zn-Cd mixtures (similar to results for Cd–Cu mixtures in Figures 5C and 5D). However, the predicted decrease in D. magna mortality as Zn concentration increased was small compared with observations (Supplemental Data, Figure S6). Finally, the calibrated model for the hatchery trout was applied to rainbow trout mortality in single-metal and metal mixture exposures of Cd, Cu, and Zn (Index V-3). Because Cu was not included in the trout model calibration, model coefficients for Cu were taken from the D. magna calibration (Table 3). Model predictions and observed mortality are compared for single-metal exposures in Figure S7 in

830

Environ Toxicol Chem 34, 2015

MMME Lessons Learned

Figure 7. Comparison of model predictions and observed 96-h mortality of hatchery rainbow trout (Oncorhynchus mykiss) and cutthroat trout (O. clarkii) in field-collected water with spiked Cd-Pb-Zn mixtures (Index 6). Mortality is shown as a function of (A) fractional coverage of total metal on the biotic ligand (uM) based on case 1 (adjusted binding coefficient [log KM] values/concentration addition [C-A]); (B) equivalent fractional coverage of total metal on the biotic ligand (uM(equiv)) based on case 2 (adjusted log KM values/independent action [I-A]); (C) the TOX function based on case 3 (adjusted metal potency factor [aM] values/ TOX addition); and (D) the equivalent TOX function (TOXequiv) based on case 4 (adjusted aM values/independent action). Observed responses (symbols) are compared with the model-predicted response curves (continuous line). Dashed lines represent plus/minus a factor of 2 in the uM, uM(equiv), TOX, or TOXequiv values at which a model-predicted response occurs. The ‘  ’ symbol represents exposure solutions that were excluded from the calibration because they were suspected of being oversaturated with PbCO3(s) (see text). See Table 1 for further description of the data.

Supplemental Data and for metal mixtures in Figure S8 in Supplemental Data. The model predicted observed mortality in Cu-only and Cd-only toxicity tests fairly well. However, the calibrated model underpredicted observed mortality in Zn-only exposures (with calculated uM and TOXi values for the observed data being roughly a factor of 2 lower than the corresponding values on the model-predicted response curve). Results for the effects of Cd-Cu-Zn ternary mixtures were consistent with findings presented previously, with observed mortality better described by independent action. LESSONS LEARNED

Biotic ligand models and the WHAM-FTOX model provide useful frameworks for predicting the effects of chemical speciation on metal bioavailability and accumulation of metals at a site of toxic action on or in an organism. The models have generally been calibrated using dissolved metal concentrations, water chemistry (pH, alkalinity, major ions, DOC), and organism effects (e.g., mortality or growth reduction). These data, however, are not sufficient to fully constrain the selection of model coefficients. This problem is compounded by the fact that model coefficients (e.g., log KM, aM, a, b) are strongly interrelated and cannot be determined independently unless measurements of metal accumulation on

toxicologically relevant binding sites (or a surrogate measure such as metal accumulation on a fish gill [18–21] or algal surfaces [22]) are also available. Although various assumptions often are employed to help constrain model calibration, these types of bioavailability models typically have more than a sufficient number of coefficients to fit toxicity data sets [3]. This makes it difficult to compare models that have been formulated and subsequently calibrated based on various underlying assumptions. In the present study, a streamlined modeling approach was therefore developed based on many of the limiting assumptions that have been employed in the previous modeling studies. The model has been applied to a series of calibration and validation data sets for P. subcapitata growth, D. magna mortality, and rainbow/cutthroat trout mortality. Based on these evaluations, we have identified the following lessons learned. Lesson 1. A single-site model is a useful and often sufficient framework for describing observed organism responses in single-metal and metal mixture exposures. Models with multiple biotic ligand sites also can be used. However, these models often add more complexity and more latitude in model calibration than is needed to describe most currently available data (which typically include only measurements of metalexposure concentrations, water chemistry, and organism effects).

K.J. Farley and J.S. Meyer

Lesson 2. Attempts to fully constrain model calibration for single-metal exposures using LFER-calculated log KM values and assuming that all metals elicit equivalent toxicities on the biotic ligand were not successful. Additional adjustments of model coefficients were generally needed. Model calibration for single-metal exposures therefore can be achieved by adjusting log KM values or by adjusting aM values (or some combination of the 2 approaches). Although both calibration strategies can provide good fits to observed responses for single-metal exposures, changing the log KM or aM values has different effects on model predictions for metal mixtures. Changing an aM value affects only the contribution of that metal to the predicted toxicity; however, changing a KM value not only affects the predicted toxicity of the metal in question, but can in principle affect the predicted toxicity of the other metal(s) via competition effects at the binding site on the biotic ligand. Lesson 3. Although different strategies can be used in model calibration for single-metal exposures, calibrating BLMs across multiple data sets and test conditions remains a difficult task. Two factors played important roles in the calibration of the streamlined model. First, characterization of the biotic ligand as a bidentate site was crucial in modeling the effects of pH on P. subcapitata growth (and probably for the effects of pH on D. magna mortality and the mortality of rainbow and cutthroat trout). Second, the Irving-Rossotti LFER for bidentate binding to negatively charged oxygen donor groups provided a good first estimate of competitive binding of metals and major cations (and subsequently other metals) to the biotic ligand. Lesson 4. Well-calibrated models for single-metal exposures are crucial in predicting metal mixture toxicity. Selection of a mixture-toxicity model (based on concentration addition or independent action) is also important. Independent action generally provided better predictions of metal mixture toxicity for evaluations in the present study (e.g., effects of Cd-Cu and Cd-Zn mixtures on D. magna mortality, effects of Cd-Pb-Zn and Cd-Cu-Zn on trout mortality), but exceptions occurred. For example, the effects of Cu-Zn mixtures on D. magna mortality appeared to be better described by concentration addition (or TOX addition). This suggests that different mixture-toxicity models (or a combination of mixture-toxicity models) may be more applicable for different mixtures of metals [23]. Lesson 5. In model calculations, a less-than-additive response was associated with competitive metal–metal interactions on the biotic ligand. This behavior only appeared to be important when the accumulation of the second metal exceeded a fractional coverage of 0.1 on the biotic ligand (e.g., computed Cu accumulations in the D. magna Cd–Cu exposure test series [Figure 6D]). The inability of the current model to adequately predict a less-than-additive response in the D. magna Cd-Zn test (Supplemental Data, Figure S6) suggests that additional adjustment of the log KM value for Zn may be necessary. Adjustment of the log KM value for major cations such as Mg2þ (which was computed to occupy most biotic ligand sites for this test series) also should be considered. In moving forward, results from model investigations should be used in guiding additional laboratory studies. Because wellcalibrated models for single-metal exposures are prerequisite in predicting metal mixture toxicity, this work should include both single-metal and metal mixture studies. For single-metal exposures, efforts should focus on the development of methods for measuring metal accumulations on site(s) of toxic action. This information will be crucial in advancing the understanding of competitive interactions and metal potency on the biotic

Environ Toxicol Chem 34, 2015

831

ligand and in helping to constrain the calibration of metalbinding coefficients. In the short-term, recent data on metal binding to fish gills [20,21] should be used, in addition to the older data [18,19] that were used to calibrate some of the HDR single-metal BLMs. Other factors, such as toxicity of Hþ and Al in low-pH waters, which has been included in the CEH model [11], also should be considered. The use of LFERs to describe metal binding to sites on or in the organism also shows promise and should receive more attention. The LFER calculations presented herein only considered sites with bidentate binding to oxygen donor groups. Extension of this approach to include both monodentate and bidentate binding as well as binding to nitrogen groups [24] and sulfur groups should be considered. Calculations of metal-carbonate and metal-hydroxide precipitation also should be incorporated into models to better evaluate metal toxicity in higher-pH waters. For metal mixtures, toxicity testing of individual metals and their mixtures should be conducted concurrently to avoid erroneous conclusions about metal mixture toxicity [25]. Otherwise, the variation associated with nonsimultaneous testing needs to be adequately addressed [26]. Models should be used as part of this effort to help identify metal mixture combinations that are more likely to demonstrate significant differences in organism response. A large part of this effort should focus on more-than-additive and less-than-additive toxicity scenarios (sensu [27]), because of their relevance in regulatory decision-making. Finally, one must remember that geochemical-based modeling approaches that are incorporated in all BLM- and WHAM-FTOX–type models may serve as a first step in explaining metal mixture toxicity. In addition to metal bioavailability and competitive interactions on binding sites on or in the organism, other factors such as metal-uptake kinetics, the movement of metals to an internal site of toxic action in the organism, and biological feedback processes (e.g., metallothionien synthesis) also may play important roles and may need to be considered in future models. SUPPLEMENTAL DATA

Tables S1. Figures S1–S8. (1.1 MB DOC). Acknowledgment—Funding for this work was provided by the Copper Alliance, International Zinc Association, Nickel Producers Environmental Research Association, and Rio Tinto. The authors are especially grateful to E. Tipping and S. Lofts (Centre for Ecology and Hydrology, United Kingdom), C. Mebane and L. Balistrieri (US Geological Survey, USA), Y. Iwasaki (Tokyo Institute of Technology, Japan), M. Kamo and W. Naito (National Institute of Advanced Industrial Science and Technology, Japan), and R. Santore and A. Ryan (HDRïHydroQual, USA) for providing their insights and thoughtful comments during the preparation of this manuscript. The authors also thank E. Van Genderen, W. Adams, R. Dwyer, E. Garman, and J. Gorsuch for sharing their insights and for coordinating the MMME project study and 2 anonymous reviewers for their careful review of the manuscript and their helpful comments and suggestions. REFERENCES 1. Van Genderen E, Adams W, Dwyer R, Garman E, Gorsuch J. 2015. Modeling and interpreting biological effects of mixtures in the environment: Introduction to the metal mixture modeling evaluation project. Environ Toxicol Chem 34:721–725 (this issue). 2. Iwasaki Y, Kamo M, Naito W. 2015. Testing an application of the biotic ligand model to predict acute toxicity of metal mixtures to rainbow trout. Environ Toxicol Chem 34:754–760 (this issue). 3. Farley KJ, Meyer JS, Balistrieri LS, De Schamphelaere KAC, Iwasaki Y, Janssen CR, Kamo M, Lofts S, Mebane CA, Naito W, Ryan AC, Santore RC, Tipping E. 2015. Metal mixture modeling evaluation

832

4. 5.

6. 7. 8. 9. 10.

11. 12.

13.

14. 15. 16. 17.

Environ Toxicol Chem 34, 2015 project: 2. Comparison of four modeling approaches. Environ Toxicol Chem 34:741–753 (this issue). Balistrieri LS, Mebane CA. 2014. Predicting the toxicity of metal mixtures. Sci Total Environ 466–467:788–799. Balistrieri LS, Mebane CA, Schmidt TS, Keller W. 2015. Expanding metal mixture toxicity models to natural stream and lake invertebrate communities. Environ Toxicol Chem 34:761–776 (this issue). Santore RC, Ryan AC. 2015. Development and application of a multimetal multibiotic ligand model for assessing aquatic toxicity of metal mixtures. Environ Toxicol Chem 34:777–787 (this issue). Tipping E, Lofts S. 2015. Testing WHAM-FTOX with laboratory toxicity data for mixtures of metals (Cu, Zn, Cd, Ag, Pb). Environ Toxicol Chem 34:788–798 (this issue). Tipping E, Lofts S, Sonke JE. 2011. Humic ion-binding model VII: A revised parameterization of cation-binding by humic substances. Environ Chem 8:225–235. Lofts S. 2012. User’s Guide to WHAM7. NERC Centre for Ecology and Hydrology, Bailrigg, Lancaster, UK. Carbonaro RF, Di Toro DM. 2007. Linear free energy relationships for metal–ligand complexation: Monodentate binding to negativelycharged oxygen donor atoms. Geochim Cosmochim Acta 71:3958– 3968. Tipping E. 2005. Modelling Al competition for heavy metal binding by dissolved organic matter in soil and surface waters of acid and neutral pH. Geoderma 127:293–304. Lofts S, Tipping E. 2011. Assessing WHAM/Model VII against field measurements of free metal ion concentrations: Model performance and the role of uncertainty in parameters and inputs. Environ Chem 8:501– 516. Carbonaro RF, Atalay YB, Di Toro DM. 2011. Linear free energy relationships for metal–ligand complexation: Bidentate binding to negatively-charged oxygen donor atoms. Geochim Cosmochim Acta 75:2499–2511. Davis TA, Volesky B, Mucci A. 2003. A review of the biochemistry of heavy metal biosorption by brown algae. Water Res 37:4311– 4330. HydroQual. 2007. The Biotic Ligand Model Windows Interface, Version 2.2.3: User’s Guide and Reference Manual. Mahwah, NJ. Kamo M, Nagai T. 2008. An application of the biotic ligand model to predict the toxic effects of metal mixtures. Environ Toxicol Chem 27:1479–1487. Niyogi S, Wood CM. 2004. Biotic ligand model, a flexible tool for developing site-specific water quality guidelines for metals. Environ Sci Technol 38:6177–6192.

MMME Lessons Learned 18. Playle RC, Dixon DG, Burnison K. 1993. Copper and cadmium binding to fish gills: Estimates of metal–gill stability constants and modelling of metal accumulation. Can J Fish Aquat Sci 50:2678–2686. 19. MacRae RK, Smith DE, Swoboda-Colberg N, Meyer JS, Bergman HL. 1999. Copper binding affinity of rainbow trout (Oncorhynchus mykiss) and brook trout (Salvelinus fontinalis) gills: Implications for assessing bioavailable metal. Environ Toxicol Chem 18:1180–1189. 20. Birceanu O, Chowdhury MJ, Gillis PL, McGeer JC, Wood CM, Wilkie MP. 2008. Modes of metal toxicity and impaired branchial ionoregulation in rainbow trout exposed to mixtures of Pb and Cd in soft water. Aquat Toxicol 89:222–231. 21. Todd AS, Brinkman S, Wolf RE, Lamothe PJ, Smith KS, Ranville JF. 2009. An enriched stable-isotope approach to determine the gill–Zn binding properties of juvenile rainbow trout (Oncorhynchus mykiss) during acute zinc exposures in hard and soft waters. Environ Toxicol Chem 28:1233–1243. 22. Slaveykova VI, Wilkinson KJ, Ceresa A, Pretsch E. 2003. Role of fulvic acid on lead bioaccumulation by Chlorella kesslerii. Environ Sci Technol 37:1114–1121. 23. Dyer S, Warne MSJ, Meyer JS, Leslie HA, Escher BI. 2011. Tissue residue approach for chemical mixtures. Integr Environ Assess Manag 7:99–115. 24. Atalay YB, Di Toro DM, Carbonaro RF. 2013. Estimation of stability constants for metal–ligand complexes containing neutral nitrogen donor atoms with applications to natural organic matter. Geochim Cosmochim Acta 122:464–477. 25. De Laender F, Janssen CR, De Schamphelaere KAC. 2009. Nonsimultaneous ecotoxicity testing of single chemicals and their mixture results in erroneous conclusions about the joint action of the mixture. Chemosphere 76:428–432. 26. Iwasaki Y, Brinkman SF. 2015. Application of generalized linear mixed model to analyze mixture toxicity: Survival of brown trout affected by copper and zinc. Environ Toxicol Chem 34:816–820 (this issue). 27. Meyer JS, Farley KJ, Garman ER. 2015. Metal mixture modeling evaluation project: 1. background. Environ Toxicol Chem 34:726–740 (this issue). 28. Meyer JS, Ranville JF, Pontasch M, Gorsuch JW, Adams WJ. 2015. Acute toxicity of binary and ternary mixtures of Cd, Cu, and Zn to Daphnia magna. Environ Toxicol Chem 34:799–808 (this issue). 29. Mebane CA, Dillon FS, Hennessy DP. 2012. Acute toxicity of cadmium, lead, zinc, and their mixtures to stream-resident fish and invertebrates. Environ Toxicol Chem 31:1334–1348. 30. Naddy R, Cohen AS, Stubblefield WA. 2015. The interactive toxicity of cadmium, copper, and zinc to Ceriodaphnia dubia and rainbow trout (Oncorhynchus mykiss). Environ Toxicol Chem 34:809–815 (this issue).

Copyright of Environmental Toxicology & Chemistry is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.

Metal mixture modeling evaluation project: 3. Lessons learned and steps forward.

A comparison of 4 metal mixture toxicity models (that were based on the biotic ligand model [BLM] and the Windermere humic aqueous model using the tox...
1MB Sizes 2 Downloads 5 Views