Using metal-ligand binding characteristics to predict metal toxicity: quantitative ion character-activity relationships (QICARs).

Ecological risk assessment can be enhanced with predictive models for metal toxicity. Modelings of published data were done under the simplifying assumption that intermetal trends in toxicity reflect relative metal-ligand complex stabilities. This idea has been invoked successfully since 1904 but has yet to be applied widely in quantitative ecotoxicology. Intermetal trends in toxicity were successfully modeled with ion characteristics reflecting metal binding to ligands for a wide range of effects. Most models were useful for predictive purposes based on an F-ratio criterion and cross-validation, but anomalous predictions did occur if speciation was ignored. In general, models for metals with the same valence (i.e., divalent metals) were better than those combining mono-, di-, and trivalent metals. The softness parameter (sigma p) and the absolute value of the log of the first hydrolysis constant ([symbol: see text] log KOH [symbol: see text]) were especially useful in model construction. Also, delta E0 contributed substantially to several of the two-variable models. In contrast, quantitative attempts to predict metal interactions in binary mixtures based on metal-ligand complex stabilities were not successful.

Quantitative structure-activity relationships (QSARs) are applied widely to predict bioactivity (e.g., toxicity or bioavailability) of organic compounds. In contrast, models relating metal ion characteristics to their bioactivity remain underexploited. A few models exist for human risk prediction [e.g., Williams et al. (I)] but quantitative models have not been fully explored for nonhuman species. This is surprising because such quantitative ion character-activity relationships (QICARs) would be extremely useful for predicting effects of untested metals during risk assessment activities. Also, qualitative ion character-activity relationships (ICARs) based on simple metal-ligand binding have been described in the literature for nearly a century. As an early example, Mathews (2) assumed that metals were most active in their ionic form (the ionic hypothesis) and correlated metal toxicity to characteristics of ion binding to biomolecules. Especially useful were characteristics reflecting bond stability with ligand groups possessing 0, N, and S donor atoms. For the last halfcentury, permutations on this approach were applied successfully by Jones (3,4), Binet (5), Loeb (6), McGuigan (7), Biesinger and Christensen (8), Jones and Vaughn (9), Kaiser (10), Williams and Turner (11), Babich et al. (12,13), Fisher (14), Newman and McCloskey (15), McCloskey et al. (16), and Tatara et al. (17,18). Modeling was often based on hard and soft acid and base theory (9,11,19). This approach has not been evaluated for its predictive usefulness despite clear indications from ICARs that QICARs were feasible. Newman and McCloskey (15) suggested that the contrasting extent of QSAR and QICAR development resulted from two factors. First, the QSAR approach was quickly incorporated into ecotoxicology because it had already proven its worth in pharmacology and human toxicology. In contrast, QICARs were not well established in pharmacology or human toxicology because the major focus of these disciplines was organic drugs and poisons. Second, chemical speciation complicates prediction because several metal species are present simultaneously and the bioavailability of each is ambiguous. However, some of this ambiguity can be removed by judiciously applying the free ion activity model (FIAM) (20). The FIAM, an extension of the ionic hypothesis, holds that the bioactivity of a dissolved metal is correlated with its free ion concentration or activity. The complication of simultaneous exposure to many species can be minimized by focusing on the free ion. Because both impediments are resolvable, no inherent obstacle impedes QICAR development to the same level of utility as that of QSARs. This paper assesses the QICAR approach for predicting metal toxicity. This is done by reanalyzing metal effects data reported elsewhere. Models are assessed by cross-validation (PRESS method as described in "Methods") relative to their effectiveness for predicting bioactivity of untested metals. An attempt is also made to extend this approach to prediction of metal interactions in binary mixtures. Abbreviations: CE50, 50% effect concentration on ability to form colonies; EC16, effective concentration for 16% response; ED50, effective dose for 50% response; 120, inhibition at the 20% level; 150, inhibition at the 50% level; LC50, concentration killing 50% of exposed individuals; NR50, neutral red response at the 50% level; TC, threshold concentration, VCF, volume concentration factor. enzyme inactivation (21,22), viability of cultured metazoan cells including cells from two fish (12,23) and a mammal (24), germination inhibition of two fungi (25), bioaccumulation in a marine diatom (14), inhibition of bacterial bioluminescence (Microtox assay, Microbics Corporation, Carlsbad, California) (16), and acute toxicity to soil nematodes (17,18). Acute toxicity was also examined for diverse aquatic invertebrates including a planarian (4), cladoceran (26), insect (1), and amphipod (27). Several data sets involved chronic exposures with lethal (1,3,8) or sublethal (8) end points. For all studies, barium toxicity was excluded from models because of its very specific interference with K+ flux in excitable tissues of metazoans (28)(29)(30)(31).

Ion Characteristics
One-and two-explanatory variable models were constructed from six ion qualities. The electronegativity (Xm) and Pauling ionic radius (r) were combined to produce a covalent index (Xm2r) reflecting the relative importance of covalent versus electrostatic interactions during metal-ligand binding (32). The ion charge (Z) and Pauling ionic radius were combined to generate a second index, the cation polarizing power (Z2Ir), reflecting the energy of the metal ion during electrostatic interaction with a ligand (32). However, no models using Z2Ir are reported here because this index did not contribute to the best candidate model for any data set. A softness index (a ) (9 11) was produced by dividing the difference between the coordinate bond energies of the metal fluoride and iodide by the coordinate bond energy of the metal fluoride. This index reflected metal ion softness, the relative tendency for the outer electron shell to deform (polarizability), and the ion's tendency to share electrons with ligands. Metal affinity to intermediate ligands such as those with 0 donor atoms was estimated with another index ((log KOHI) based on the first hydrolysis constant of the ion, i.e., KOH for Mn+ + H20 -* MOHn-1 + H+ (15). Following the approach of Kaiser (10), AN/AIP and AEO were also explored in model development. Log AN/AIP did not improve models, as suggested by Kaiser (10); AN/AIP was used instead. Atomic number (AN), notionally reflecting ion inertia or size, was combined with AIP (the difference in ionization potentials for the ion oxidation number OX and OX-1), which reflected ionization potential. The absolute difference between the electrochemical potential of the ion and its first stable reduced state (AE0) reflected an ion's ability to change electronic state. Values for these ion characteristics used in this study are tabulated by McCloskey et al. (16). Model Assessment Linear regression models were generated with these six variables (Xm2r, Z21r, ap, Environmental Health Perspectives a Vol 106, Supplement 6 * December 1998 llog KoHl, AN/AIP, AE0), and the SAS Procedure GLM general linearized model (33). Models including Z2Ir were not reported for reasons already stated. Three levels of model selection followed model generation. This procedure was applied to models including divalent metals alone or all metals regardless of valence. First, the contribution of a variable to each model was tested for statistical significance (Fstatistic from Type III sum of squares, a = 0.05). Only models in which all variables contributed significantly were considered further. Second, the predictive potential of models was estimated with an F-ratio approach because usefulness for prediction is not reflected accurately by a model's statistical significance. More rigorous criteria must be applied. A ratio of the observed F statistic (regression sum of squares divided by the residual sum of squares) to the critical F statistic (a = 0.05) greater than 4 to 5 is one accepted, albeit arbitrary, threshold for acceptable predictive utility (34). The most stringent Fobserved/Fcritical of . 5 was adopted here.
Finally, if more than one useful model existed for a data set, the best was selected by minimum Akaike's information criterion estimation (MAICE) (35). With MAICE, models that differ in complexity (i.e., one-vs two-explanatory variables) can be compared. An Akaike's information criterion (AIC) was calculated with the log likelihood function of each model [details can be found in Yamaoka et al. (35) and Newman and McCloskey (15)]: AIC = -2(log likelihood) + 2P, where P is the number of estimated parameters in the candidate model. The model with the smallest AIC was judged to contain the most information. With this three-step procedure, the best model was selected from among those that were potentially useful for prediction.
Cross-validation was performed on the best divalent metal models to estimate the magnitude of deviations in effect prediction for unknown metals. For each of the 13 divalent metal data sets producing potentially useful models, a series of models was generated after omitting one metal at a time. Each time this was done, the ion characteristics of the omitted metal were placed into the model to predict an effect for the omitted metal. This cross-validation (36) was done with the option PRESS in SAS Procedure REG (33). The deviation from perfect prediction was expressed as the percentage [(observed effectmetal i -predicted effectmodel without metal j)/observed effectmetal , X 100. Median and interquartile ranges for these percentages summarize the general deviations from perfect prediction. Interacdons in Binary Mixtures Bacterial bioluminescence data for binary mixtures of metals (15) were examined statistically to assess the hypothesis that metals with strong and similar covalent binding to ligands will interact strongly. The qualitative conclusions of Newman and McCloskey (15) were tested statistically for two series of mixtures: Cu in combination with Ca, Cd, Hg, Mg, Mn, Ni, Pb, or Zn; and Mg in combination with Ca, Cd, Hg, Mn, Ni, Pb, or Zn. Interactions were assessed statistically using the SAS Procedure MIXTURE (33) with an interaction term (i.e., metal, xmetal2).
Statistical significance and magnitude of the interaction terms were used to assess interactions between paired metals. If the above hypothesis was correct, the intensity of interaction would be greatest between Cu and similar metals (e.g., Hg or Cd) and lowest for Cu and dissimilar metals (e.g., Mg or Ca). In the second series of mixtures there would be little interaction between Mg and other metals.

Results
Models for Dalent Metals High correlation coefficients were associated with many one-and two-variable models for divalent metal effects ( Figure 1 and Table  2). Based on a stringent criterion of an F ratio. 5, 13 of the 19 data sets had at least one model of predictive utility. Data sets failing to produce useful models involved cultured cell viability (three studies), in vitro inactivation of carbonic anhydrase, and inhibition of fungal germination. Fruit fly mortality data also failed to produce a useful model according to our stringent criterion but nevertheless had a high F ratio of 4.5.
Oneor two-variable models of most predictive promise included llog KOHI or 6p.
Several two-variable models, especially those including AE0, were also among those with predictive promise. The covalent index (Xm2r) alone or combined with another variable never produced the best model for any data set. Five of the thirteen most informative models (MAICE) were single-variable models. The median r2 for these best and predictively useful models was 0.90 (range 0.78 to 0.97). Approximately 90% of the variation in metal effect could be explained by the models.  Cross-validation of the best divalent metal models (Table 3) indicated that the median deviations between observed and predicted effects were small. The median deviations were less than 22%; most were closer to 10%. For comparison, a wellknown QSAR model for bioconcentration of eight organic compounds in fish (37) had a median difference of -7% and an interquartile range of -13 to 14%. However, many models poorly predicted effects for specific metals. These metals tended to be extreme class a (e.g., Mg), class b metals that undergo considerable speciation in solution (e.g., Hg, Pb), or metals with the tendency to precipitate from solution (e.g., Mn). Under the assumption that speciation contributed to some of these poor predictions by models built from total metal concentrations, concentrations of free metal ion were estimated with the MINTEQA2 Version 3.10 program (38) for two data sets involving bacterial bioluminescence (15,16). These assays were conducted in contrasting media having speciation similar to marine (15) or freshwater (16) environments. Except for Hg in media having speciation similar to marine systems, EC50 values of metals were expressed as the free ion concentration. The EC50 for Hg was expressed in terms of the free ion plus neutral chloro complex concentration because neutral chloro complexes of Hg can also be bioavailable (39). In both media the extremely discordant predictions were greatly diminished or eliminated if EC50 values were expressed in terms of the speciated metal concentration (Table 3).

Models Induding All Metals
Although correlation coefficients were lower than those for the divalent metal models (median 0.80, range 0.67 to 0.87), useful models including all metals (mono-, di-, and trivalent) were generated for 13 of the 19 data sets (Table 4). Approximately 80% of the variation in effect for metals could be attributed to the explanatory variables. Eight of thirteen data sets producing useful models had the best (MAICE) model involving only one explanatory variable.
Again, the llog KOHI or sp indices contributed to many of the best one-variable models. As with the divalent metal models, data sets failing to produce useful models tended to be those for in vitro enzyme inactivation or cultured cell viability. Data for bioaccumulation of metals in a marine diatom and crustacean toxicity also failed to produce useful models.

Metal Mixture
Although there were qualitative indications of concentration-dependent interactions between metals with similar and high covalent binding tendencies, no such trends were noted in this formal analysis. The only significant trends in the intensity of the interaction term for both series of mixtures was a consequence of increasing LC50 values with decreasing covalent interactions; this trend was an artifact of the data structure. Regardless of whether Cu or Mg was combined with metals, there was a upward trend in the interaction term, with increasing tendency of the competing metal to interact covalently with ligands. Such a trend for the Mg series of binary mixtures was inconsistent with predictions from the initial hypothesis. The results did not support the initial hypothesis that mixture interactions could be predicted from the tendency to covalently bind with ligands.

Conclusion
Quantitative ion character-activity relationships can be developed for a range of effects based on metal-ligand binding theory. Estimations of speciation and application of the FIAM were not required to develop useful QICARs for some metals in the data sets. Our work with QICAR development for microbial bioluminscence (15,16) and nematode toxicity (17,18) supports this observation. However, there are clear indications that calculation of free ion concentrations or activities will greatly improve modeling, i.e., eliminate or reduce the magnitude of anomalous predictions for some class b metals.
The results for the relatively simple in vitro enzyme inactivation and cultured cell viability studies illustrate the difficulties associated with using models based on total metal concentration. These data sets involved buffered or complex media, i.e., the enzyme inactivation in a buffered phosphate solution and cell culture experiments in complex media containing components such as 10% fetal calf serum. Another data set failing to produce a useful model was associated with a high ionic strength media, i.e., bioaccumulation in a marine diatom. This association of model failure with complex exposure media suggests that speciation calculation would improve modeling because speciation is most extensive under these conditions. This is further supported by the diminution of deviations from perfect prediction during cross-validation of models considering speciation. If speciation is ignored, predictions of effect should be done cautiously for class b and some intermediate metals characterized by extensive speciation in solution. Published trends for metal speciation in marine (40) and freshwater (40,41) systems can be used to identify those metals for which speciation should be considered during QICAR development. The ion characteristics of most general value in constructing QICARs were llog KOHI and cp, although other variables such as AEO were also important in several models. The llog KOHI reflects the tendency for a metal ion to form a stable complex with intermediate ligands.
Intermediate ligands on biomolecules would include groups with 0 donor atoms (e.g., carboxyl groups). This suggested that binding with such functional groups is important in determining the relative bioactivity of metals. The softness index (ap) quantifies the ability of a metal ion to accept an electron during interaction with a ligand. It reflects the importance of covalent interactions relative to electrostatic interactions (32) in determining intermetal trends in bioactivity.
The results also suggest that QICARs based on the characteristics used in this study are best developed for metals of similar charge. Although models based on variables such as ap did produce viable models, Ahrland (42) and Williams and Turner (11) argue against the application of op for metals differing in charge. Instead, variables adjusting for differences in charge, such as ck (11), may be required.
Effective application of the QICAR approach may also be improved by careful examination of the values used to generate the explanatory variables. Considerable judgment is required when selecting among published estimates. More involved analysis of these data for application to QICAR generation is currently required. Regardless, QICARs are now feasible, especially if they were produced with speciation concentrations for metals of similar charge.
Ecological risk assessment would be enhanced by reliable models for predicting effects of untested metals from known effects of tested metals. In the absence of complete information on the effect of all metals of concern on each important species under a variety of conditions, the ability to interpolate from existing data to predict effects for untested metals would improve the accuracy of assessments. QICARs would be particularly useful in preliminary screening and in situations analogous to those in which QSARs are currently applied. Our results suggest that the QICAR approach would be extremely useful for this purpose. However, several resolvable issues require attention before the QICAR approach has the same general usefulness as the QSAR approach. These issues include exploration of more explanatory variables, careful evaluation of ionic qualities used to calculate explanatory variables, examination of models capable of predicting effects for widely differing metals (e.g., metals of different valence states), effective inclusion of chemical speciation, examination of more effects, and assessment of the applicability of QICARs to phases such as sediments, soils, and food.