ToxSci Advance Access originally published online on December 7, 2007
Toxicological Sciences 2008 102(1):187-195; doi:10.1093/toxsci/kfm294
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Published by Oxford University Press 2007.
Modeling and Assaying Dioxin-Like Biological Effects for both Dioxin-Like and Certain Non-Dioxin–Like Compounds
* Division of Systems Toxicology
Division of Genetic and Reproductive Toxicology
Dioxin Group, Arkansas Regional Laboratory
Department of Biostatistics, University of Arkansas for Medical Sciences, Little Rock, AR 72205-7199
1 To whom correspondence should be addressed at Building 26, HFT-233, National Center for Toxicological Research, 3900 NCTR Road, Jefferson, AR 72079. Fax: (870) 543-7686. E-mail: jon.wilkes{at}fda.hhs.gov.
Received August 16, 2007; accepted December 5, 2007
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
13C NMR data have been correlated to Toxic Equivalency Factors (TEFs) of the 29 PCDDs, PCDFs, or PCBs for which non-zero TEFs have been defined. Such correlations are called quantitative spectrometric data-activity relationship (QSDAR) models. An improved QSDAR model predicted TEFs of 0.037 and 0.004, respectively, for 1,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) and 1,2,3,4,7-pentachlorodibenzo-p-dioxin (PeCDD), both among the 390 congeners for which zero value TEFs are assumed. A QSDAR model of Relative Potency (REP) values estimated the corresponding values as 0.115 and 0.020. Results from both models indicated that these two congeners may exhibit significant dioxin-like toxicity. If other such congeners have non-zero toxicity, TEF-based risk assessments of some dioxin-, furan-, or PCB-contaminated sites or foods may underestimate toxicity. Both models were extensively cross-validated and the TEF model was externally validated. We confirmed the predictions by an independent in vitro method, a luciferase gene expression assay based on mouse liver cells that found REPs of 0.027 and 0.013, respectively, for 1,3,7,8-TCDD and 1,2,3,4,7-PeCDD. The QSDAR-estimated and gene-expression assayed values agreed. The models were used to predict activity for an applicability domain including 108 non-2,3,7,8 dioxin, furan, or PCB congeners and 2,3,7,8-tetrachlorophenothiazine, a dioxin analog proposed as a drug candidate. This study showed that QSDAR prediction followed by a relatively inexpensive in vitro assay could be used to nominate a few candidates among hundreds for further investigation. It suggested that in silico and in vitro nomination protocols may facilitate practical risk assessment when chemical family members exhibit different degrees of toxicity operating via a common mechanism.
Key Words: regulatory/policy; dioxin; polychlorinated biphenyls; QSAR; biological modeling.
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
2,3,7,8-Tetrachlorodibenzo-p-dioxin (TCDD) and 1,2,3,7,8-pentachloro dibenzo-p-dioxin (PeCDD) are the most biologically active of the 419 polychlorinated dibenzodioxin (PCDD), polychlorinated dibenzofuran (PCDF), and polychlorinated biphenyl (PCB) congeners. These 419 compounds cause to differing degrees, depending on their chlorine substitution pattern: a wasting syndrome; immuno- and neurotoxicities; carcinogenicity; adverse reproductive, developmental, and endocrine effects; chloracne; hepatotoxicity; the induction of drug-metabolizing enzymes; and induction and suppression of genes (Safe, 1997/1998
; van den Berg et al., 1998).
Toxicities of a 29 member subset of the 419 congeners are expressed relative to that of 2,3,7,8-TCDD (toxic equivalency factor, TEF 
1.0). The TEF is an assigned number based on an interim concept of PCDD toxicities that assumes additivity of components in PCDD, PCDF, and PCB mixtures. The concept is appropriately used in some mixture applications (Stahl et al., 1992) but not in others (Safe, 1997/1998). TEFs have been defined in consideration of toxicological and biological end points such as body weight loss, tumor promotion indexes, developmental and reproductive toxicity, dissociation constants for aryl hydrocarbon receptor (AhR) binding, mean lethal dose (LD50), mean effective dose (ED50), and medium effective concentration (EC50). The validity of this scale of relative toxicities has been demonstrated in vivo and in vitro (Birnbaum and DeVito, 1995; Safe, 1994, 2002). Despite its limitations (Safe, 1998), the TEF concept is used as the standard measure of untoward biological effects of dioxin-like compounds, especially when assessing toxicity of congener mixtures typical of environmental or food contamination.
Operating under the twin assumptions that toxic activity resides in the 2,3,7,8-chlorine substitution configuration and that the AhR is mechanistically involved, an expert committee of the World Health Organization has assigned nonzero TEF values to 29 of the 419 congeners (van den Berg et al., 1998, 2006). Besides aryl hydrocarbon binding, factors such as environmental persistence, pharmacokinetics, and pharmacodynamics were considered in assigning nonzero TEFs to these 29 congeners.
TEFs are semiquantitative, expressed to the nearest order or half order of magnitude with "half" being defined on the log scale. Uncertainty remains that congeners assigned as zero TEFs lack dioxin-like toxicity. Early results suggest that congeners not regarded as dioxin-like because of their rapid detoxification and elimination in high dose experiments may not be so rapidly eliminated under low dose exposures (van den Berg et al., 1983).
Toxicity modeling of PCDDs, PCDFs, and PCBs is desirable: (1) because of the potential biological and environmental toxicities of dioxin-like compounds; (2) because of the absence for most congeners of biological assay data with which to estimate the toxicities; and (3) because comparisons among the tested congeners show that minor differences in chlorine substitution patterns and/or carbon backbone structure lead to major differences in toxicity. The absence of toxicity testing of each possible congener prevents confident assignment of zero TEFs to the 390 congeners currently assumed, for risk assessment purposes, to have insignificant toxicity. This situation suggested the potential value in development of models to predict TEFs.
Various molecular modeling systems have been developed to predict congener toxicity, examples of which are reported below. Most models attempt to predict binding affinity or one of the other properties correlated to toxicity but they do not typically attempt prediction of TEFs. It may be that the clearly approximate nature of TEFs has discouraged attempts to model them quantitatively. Our own studies have shown a 90% correlation between nonzero TEF values and corresponding AhR binding values (Buzatu et al., 2004). This suggested that TEFs, like binding affinities, could be modeled. The TEF modeling performed in this work did not make use of binding affinities alone but depended upon the entire constellation of factors used when the nonzero TEFs were defined. Besides modeling TEFs, we also constructed a relative potency (REP) model of the same compounds. The REP model tracked AhR binding only.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
QSDAR general.
Quantitative spectrometric data-activity relationship (QSDAR) models are computational platforms developed by correlating patterns in analytical spectra with a quantitative expression of a biological characteristic. Here the QSDAR models correlated each training set compound's predicted 13C nuclear magnetic resonance (NMR) spectrum with the logarithm of its TEF or REP. The training set compounds were the 29 "dioxin-like" (nonzero TEF) polychlorinated dioxins, furans, or PCBs. These models could then be used to test an applicability domain including many other polychlorinated dioxins, furans, and PCBs; to estimate TEFs or REPs for each domain member (or other similar compounds) and to flag those which might make significant contributions to the toxicity of mixtures. Our previous reports demonstrated that QSDAR provided a rapid and simple way to model the AhR binding of dioxin-like molecules (Beger and Wilkes, 2001
; Beger et al., 2002). The most reliable AhR binding model obtained an r2 = 0.95 and leave-one-out (LOO) q2 = 0.88 (Beger et al., 2002). Other methods for producing relative activity models have been reported and a discussion of these along with a comparison to QSDAR is provided in Supplementary Data.
The same QSDAR models that succeeded in modeling toxicological properties of nonzero TEF congeners also predicted that several congeners currently assigned zero TEFs might be toxic (Buzatu et al., 2004). In this work we were particularly interested in two for which we had purified samples and an in vitro method (discussed in the next section) to assess independently the sample toxicity: 1,3,7,8-TCDD and 1,2,3,4,7-PeCDD. Because such congeners have a zero TEF, any analytically determined concentrations in an environmental mixture when multiplied by the zero TEFs would yield zero as their contribution toward the estimated toxicity of the mixture. In mass spectrometric analyses of environmental and food samples for quantitative determination of dioxins and furans, peaks of zero TEF congeners consistently appear and are ignored. (See "HRGC_HRMS Single Ion Chromatograms of two seafood samples" provided in Supplementary Data.) Having been provided highly purified samples of 1,3,7,8-TCDD and 1,2,3,4,7-PeCDD, we tested them for toxicity, and based on the assay results calculated experimental REPs for comparison to corresponding QSDAR-predicted TEF and REP model predictions.
QSDAR modeling specifics.
NMR spectra for each of the 29 nonzero and the two zero-TEF dioxin, furan, and PCB congeners were generated using CNMR version 10.0 (ACD Labs, Toronto, Ontario, Canada). The 220 ppm range of the 13C NMR spectra was divided into 1 ppm wide bins and the intensities (occupancies) of predicted NMR peaks for each congener were used to populate the appropriate bins. Every chemical shift from 124.00 to 124.99 was placed into bin 124 with a value of 100 for each atom that generated it; two chemical shifts in the same spectrum that occur between 124.00 and 124.99 would produce a bin population of at least 200, etc.
Because these dioxins, furans, and PCBs contained no aliphatic carbon atoms, the pertinent 13C NMR chemical shift range was effectively reduced from 220 to 39 bins. Analysis was further limited to bins populated in two or more spectra so that the model's inferences would not be based on single occupancy phenomena, which could lead to model "memorization" having little or no predictive value. Validation of and predictions from the model can yield good results if and only if the model has not memorized data in its training set (has not over-trained) but has generalized from the patterns.
Several pattern recognition techniques have been used successfully to build QSDAR models. These include artificial neural networks, principal component and canonical variate analysis, and multiple linear regression (MLR) (Beger and Wilkes, 2001; Beger et al., 2002; Buzatu et al., 2004). MLR provided by Statistica Version 7.0 (StatSoft, Tulsa, OK) was used in this work.
MLR forward regression added bins to a model in sequence starting with the bin showing the most statistical significance in relation to the endpoint, either the assigned TEF or the experimental REP. The regression first calculated a single β coefficient which, when multiplied by each congener's occupancy for the selected bin, minimized the sum of the squared deviations from each corresponding known TEF or REP used to build the model. The overall quality of the model was assessed in part by an F-score (a statistical figure of merit that increases with increasing quality). The second most significant and succeeding bins were added and optimal coefficients, βi, one for each bin, were calculated. We adopted the rule that, to justify adding another bin, its addition had to increase the F-score by at least 1.0 compared with the previous model. Also, regardless of the effect on the F-score, we would not add a bin that did not pass the conventional standard of statistical significance (p < 0.05).
To quantify a model's internal consistency, the program calculated a coefficient of determination (r2) and for its predictive value under LOO cross-validation, a q2 statistic. The q2 was calculated exactly as r2 except that the estimated TEF or REP for each congener was obtained from a model built without using that congener. The q2 in this case was based on predictions from 29 models, each missing one of the 29 congeners. As with r2, we compared the 29 predictions to their corresponding real values and calculated the statistical accuracy. Because predictions were poorer when each model had not trained on the specific spectrum for which the prediction was made, the overall predictive quality, q2, was smaller than the corresponding r2. If there was not much falloff from r2 to q2, this indicated that the model containing "i" bins and based on spectra from all 29 congeners was useful for predictive purposes. So long as increasing "i" (the number of bins) yielded increasing q2, the model's predictive ability was improving and significant memorization was not occurring.
We also conducted a number of more rigorous leave-three-out (L3O) cross-validation and external validation exercises. For both TEfs and REPs, 10 models of the training set congeners, each depleted by three, were built with the omitted congeners then being predicted. In each of the 10 models different combinations of congeners were omitted. For each model, r2 and q32 values were obtained as was an overall qtest2, the latter reflecting the correlation between true values and those predicted for all left out congeners in all the cross-validations. These cross-validation models were built using the bins discovered from the non-depleted training set, so although each model had not been trained on the particular left out spectra, it did contain a residual influence from them with respect to bin selection. For the TEFs, we also built 10 L3O models, in each case defining a new set of optimal bins. There were no contributions at all from omitted spectra. These 10 models provided external validation of the TEF model.
CALUX general.
The Chemically Activated LUciferase gene eXpression (CALUX) assay yields AhR binding data that can be used to estimate REP. Estimation of dioxin-like toxicity in mixtures has taken several different approaches which are not equivalent in their precise meaning but give somewhat comparable results: (1) high-resolution gas chromatography (GC)–high-resolution mass spectrometry(MS) (with toxicity inferred from the sum of products from each congener's concentration and TEF) is the international standard analysis but is time consuming, costly, complex, and requires highly-trained analysts; (2) immunological methods are comparatively simple, but the antibody is relatively new and a database has not been established (Sugawara et al., 1998); (3) the CALUX assay is based on the biological end point AhR binding as well as induction of CYP1A1 and can be used to establish REP, giving values similar to toxic equivalency for most chlorinated dioxin-like compounds (Brown et al., 2001). The CALUX assay has been used successfully in numerous studies in a variety of applications to evaluate the potency of dioxin-like molecules (Covaci et al., 2002; Hamers et al., 2002; Tsutsumi et al., 2003; van Overmeire et al., 2001) and has been validated for these types of applications (Hooper et al., 2000; Koppen et al., 2001; van Overmeire et al., 2000).
QSDAR modeling specifics.
NMR spectra were generated using CNMR version 10.0 (ACD Labs) for each of the 29 nonzero and 108 zero TEF PCDD, PCDF, and PCB congeners in the models’ estimated applicability domain (for definition of this domain see the Results section below) and for a drug candidate dioxin analog. The 220 ppm range of the 13C NMR spectra was divided into 1 ppm wide bins and the intensities (occupancies) of predicted NMR peaks for each congener were used to populate the appropriate bins.
MLR pattern recognition provided by Statistica Version 7.0 (StatSoft) was used in this work. The same modeling technique has been used before to assess dioxin biological activity (Beger and Wilkes, 2001; Beger et al., 2002; Buzatu et al., 2004).
The number of bins used to build TEF or REP models was chosen by the objective, statistically valid process described above. An 11-bin model of the 29 nonzero TEF congeners met all criteria, had excellent r2 and q2 values, and was used to predict TEFs for the 108 dioxin-like congeners in the applicability domain, all of which are assigned zero value under the World Helth Organization (WHO) system, including the two for which CALUX values were obtained. A 10-bin model of the same 29 congeners was also created in which the toxicological end points were REP values. Values used were those published in the 50th percentile of in vitro REPs, of the REP2005 database (Hawes et al., 2006). An example of typical spectral input for such models is provided in Supplementary Data, "Binned spectra 29 TEFs and 2 unknowns." The REP2005 model was also used to make predictions for each of the 108 members in the applicability domain.
Purity of the two test samples.
The two test compounds were dissolved in hexane and assayed for purity by high-resolution GC/high-resolution MS. The results showed very high purity with a few small extraneous peaks. The 1,3,7,8-TCDD chromatogram showed two additional peaks, one of which could have been 1,2,3,7,8,-PeCDF (TEF = 0.05). Their area was very small relative to the major component so their potential contribution to toxic equivalency (one ten-thousandth that of the 1,3,7,8-TCDD) was negligible. The 1,2,3,4,7-PeCDD analysis among the various single ion traces showed four additional peaks, one of which might have been 2,3,7,8-TCDF (TEF = 0.1; REP = 0.067) and another, 1,2,3,4,7,8-hexachloro dibenzo-p-dioxin (TEF = 0.1; REP = 0.075). REPs for these congeners were from Behnisch et al. (2003) and Windal et al. (2005). Assuming that these impurity peak identities were correct and not some other tetra- or hexa-chlorinated furan or dioxin, we used the corresponding REPs to estimate the maximum extra toxic contribution. This reduced by 7.5% the value reported for our CALUX assay of the 1,2,3,4,7 penta congener. The assumption was conservative in the sense that if the identifications were incorrect and these peaks were from isobaric congeners with a different substitution pattern and no toxicity, then they would by convention have zero TEFs and the CALUX estimate for this congener should have been up to 7.5% higher. The toxicities of the two test congeners were not overestimated.
CALUX assay specifics.
The mechanism of action by which CALUX detects the planar polychlorinated aromatic hydrocarbons has been elucidated (Murk et al., 1996). A summary of the process is provided in Supplementary Data, "CALUX Mechanism of Action and Experimental Details." It depends on transfection of a mouse or rat liver cell line with a luciferase reporter gene that gives luminescence whenever the cytoplasmic AhR is bound by a dioxin-like compound. That is, CALUX tracks AhR binding and induction of Cyp1A1.
Solutions of 1,3,7,8-TCDD and 1,2,3,4,7-PeCDD were prepared in dimethyl sulfoxide (DMSO) to a final concentration of 32 and 54 ng/ml, respectively. Prior to dosing, 4 µl of the DMSO solutions were suspended in 400 µl of supplemented medium; 190 µl of this medium were then used as the dosing solution. To determine a standard curve, 2,3,7,8-TCDD (Wellington, Guelph, Ontario, Canada) was used at concentrations of 100, 50, 25, 12.5, 6.25, 3.13, 1.56, 0.78, 0.39, 0.20, and 0.10 pg/well. Plating for the two test samples and for the standard curves was performed in five different assays. Following a 24-h incubation period to allow optimal luciferase gene expression, the medium was aspirated, and the induction of luciferase was quantified using the luciferase assay kit (Promega, Madison, WI) with a luminometer (Berthod Detection System, Oak Ridge, TN).
Calculation of REPs from CALUX binding.
Data for the dose response of 2,3,7,8-TCDD were fitted to a sigmoid curve described by the Hill equation using a least squares algorithm. The luminescence value for each test compound response was compared with that from 2,3,7,8-TCDD (TEF 1.0; REP 1.0), and REPs for each dilution and from each of the five replicate assays were calculated, with results averaged to determine the REP and a standard deviation. Based on the purity assay results, the gross REP for 1,2,3,4,7-PeCDD was reduced 7.5% as previously described to give its net REP.
Predictions for a tetrachlorophenothiazine, and 108 zero-TEF dioxins, furans, and PCBs.
We submitted to the TEF and REP models the predicted spectrum for 2,3,7,8-tetrachlorophenothiazine, a dioxin analog. The structure is identical to 2,3,7,8-TCDD except that the phenyl ring-joining oxygen atoms have been replaced by one each of –S– and –NH–. This compound is being tested as a new drug candidate potentially useful because it has much of 2,3,7,8-TCDD's affinity for AhR but is very rapidly metabolized (Fried et al., 2007). Accordingly, we expected that our QSDAR REP2005 model, which considered only AhR binding, would predict a high value whereas our QSDAR TEF model, which included sensitivity to other factors like ease of biodegradation, would predict a lower value.
We submitted predicted spectra for 108 other zero-TEF congeners to the models. The accuracy of these predictions could not be confirmed, even by CALUX, because no standards were available. It was possible to assess the consistency of predictions for a congener by comparing results from the TEF and REP2005 models.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
Results of the CALUX experiments and QSDAR models are summarized in Table 1 and Figures 1 and 2. The CALUX assay attributed positive biological activity to both 1,3,7,8-TCDD and 1,2,3,4,7-PeCDD. This has been reported by others (Holcomb et al., 1998
; Mason et al., 1986).
|
|
|
The 11-bin log(TEF) model used chemical shift bins 111, 117, 124, 125, 130, 132, 134,135, 138, 142, and 145 ppm from the 13C NMR spectrum. A discussion of the apparent biological significance that correlates to each of these NMR bins, is provided in Supplementary Data, "Discussion of Model Results, Details."
Figure 1 plots the logarithm of WHO-defined TEFs versus the log of their corresponding QSDAR predictions using the 11-bin MLR TEF model that had an r2 of 0.96 and q2 of 0.93. The most toxic congener—2,3,7,8-TCDD—has TEF 1.0, logarithm = 0, and so appears near the upper right corner of the plot. The variation between defined and predicted TEF values is attributable, in part, to the semiquantitative nature of the defined values being modeled. TEF predictions for the two CALUX-assayed and other test compounds were based on this model.
Figure 2 plots the logarithm of the 50% REP2005 values from Hawes et al. (2006) versus their corresponding QSDAR model predictions for the 29 dioxin-like congeners. Again, the most toxic congener, 2,3,7,8-TCDD, has REP 1.0, logarithm = 0, and appears near the upper right corner of the plot. The variation between QSDAR predictions and the REP2005 values is attributable, in part, to the fact that each database value is the mean of a collection from a variety of studies using different in vitro methodologies. Not all of the studies contained acceptable data for every congener. Therefore, unlike the REPs used for the sample purity compensation calculations, the database means do not represent quantities derived from a single, consistently defined comparison set. We constructed a 10-bin MLR model having an r2 of 0.92 and q2 of 0.88. REP predictions for the two CALUX-assayed and other compounds were based on this model. Details about bin significance are also found in Supplementary Data, "Discussion of Model Results, Details."
Results for the cross-validation and external validation exercises were good. The TEF and REP2005 values for L3O cross-validation (using the original 11 or 10 bins) were qtest2 = 0.80 and 0.72, respectively. The 10 L3O external validation models, each based on six or seven bins rather than the 11 used in the original TEF model, produced a qtest2 of 0.63. Details of these cross-validation models are found in Supplementary Data: "TEF CV 10 percent averages and STDs" and "REP CV 10 percent averages and STDs". Bins used in the various external validation models are listed with other details Supplementary Data: "TEF 10 Percent External Validation Tests."
Table 1 compares the QSDAR TEF and REP predictions plus the experimentally determined CALUX REPs with REPs calculated from AhR binding data determined by another laboratory (Mason and Safe, 1986). The CALUX REPs agree with relative toxicity inferred from the AhR binding results of Mason and Safe (1986) having error factors of only 2.1 for 1,3,7,8-TCDD and 8.0 for 1,2,3,4,7-PeCDD. It is not surprising that the CALUX assay, which measures a downstream effect of AhR binding, agrees with AhR binding itself. CALUX REPs have also been found to correspond closely to WHO TEFs (Behnisch et al, 2003; Windal et al., 2005).
The TEF and REP models predicted the activity of 2,3,7,8-tetrachlorophenothiazine as REP = 0.024 and TEF = 0.00009 whereas an experimental REP (based on CyP1A1 induction, not AhR binding) of 0.003 has been reported (Fried et al., 2007). As expected for this compound, the REP model predicts high (by a factor of 8 compared with in vitro experimental) whereas the QSDAR TEF model, which implicitly encompasses metabolic detoxification factors, predicts very low (by a factor of 33). 2,3,7,8-Tetrachlorophenothiazine lies outside the estimated applicability domain of these models and yet predictions were reasonable.
The estimated applicability domain for both the TEF and REP2005 models was polychlorinated dioxins, furans, and PCBs containing four or more chlorine substituents with a few other constraints:
- (1) Dioxins and furans had to be chlorine substituted at three of the 2,3,7, or 8 carbon positions; the fourth or more chlorines would necessarily be substituted elsewhere for zero-TEF congeners.
- (2) PCBs had to possess at least 3 chlorines attached to carbons in the 3,3',4,4',5, or 5' positions; at least one of these positions on each ring had to be chlorine substituted. PCBs were excluded when three of the 2,2',6, or 6' carbon positions had an attached chlorine because such PCBs are not coplanar and cannot interact with the AhR.
- (2) PCBs had to possess at least 3 chlorines attached to carbons in the 3,3',4,4',5, or 5' positions; at least one of these positions on each ring had to be chlorine substituted. PCBs were excluded when three of the 2,2',6, or 6' carbon positions had an attached chlorine because such PCBs are not coplanar and cannot interact with the AhR.
Results for the dioxins and furans meeting these criteria seemed reasonable in most cases. This applicability domain definition allowed predictions for 12 zero-TEF PCDDs, 30 zero-TEF PCDFs, and 66 zero-TEF PCBs. These predictions are supplied with comments in Supplementary Data, "Unknown TEF and REP Predictions." Of the 12 PCDDs, six showed predicted TEFs and REPs 
0.01 and could be considered for CALUX assay as part of a congener nomination protocol. Of the 30 PCDFs, 15 met the same criteria for potential concern. Of the 66 PCBs, 32 had predicted TEFs and REPs 0.0001, a lower cutoff that seemed appropriate because the PCBs are typically found in environmental contaminations and food at much higher concentrations than dioxins and furans. Obviously the number of congeners that might be of concern varied greatly depending on the cutoff used. For PCBs, raising the cutoff from 0.0001 to 0.001 reduced the number "of concern" from 32 to only 9. In the file "Unknown TEF and REP Predictions," the congeners of concern are distinguished from others using orange fonts. We are not arguing for the particular cutoff values used here; rather, we are demonstrating inference derivable from accurate toxicity models.
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
Our results demonstrate the ability of QSDAR to make plausible estimates of either semiquantitative TEFs or somewhat ambiguously defined REPs for dioxin and dioxin-like compounds. The estimates’ accuracy is even more impressive considering that the training set used for model development did not include any congener examples lacking the minimal 2,3,7,8 chlorination substitution pattern. That is, the two models in this case were being consulted to estimate values for compounds somewhat beyond the range of structural variability in the training set. Because only compounds having nonzero TEFs were used to create the QSDAR model, and the only nonzero TEFs always included at least four chlorines substituted at the 2,3,7, and 8 positions, both compounds predicted and CALUX-assayed here lay outside the models’ range of variability but, we contend, not outside the model's applicability domain.
The 11 bins used for the TEF model were selected based on the objectively defined procedure described in "Materials and Methods." The actual bins are listed with more discussion in Supplementary Data: "Discussion of Model Results." It is encouraging that many statistically chosen bins were related to structural features known to be critical in determining dioxin-like activity. There were bins associated exclusively with dioxins or furans, and one bin was included (at 138 ppm) that represented adjacent nonchlorinated positions in a PCB. There were multiple bins associated with chlorination of phenyl rings.
It is probably surprising to students of dioxin toxicity that bins at 135 and 142 ppm were selected for inclusion in both the TEF and REP2005 models. Carbons attached to oxygen atoms in furans and dioxins cannot be chlorinated and would not be expected to convey information relative to the 2,3,7, or 8 positions most strongly associated with dioxin-like toxicity. However, because NMR chemical shift values are affected by substituents through as many as five bonds and with particular strength when the bonds belong to aromatic systems, the carbons attached to oxygen atoms are able to integrate the effects of multiple, remote substitutions at the 2,3,7, and/or 8 positions. The analogous situation holds for the nonchlorinated PCB carbons represented at 138 ppm. These two 1, 1' atoms bond the two phenyl rings together, which explains why they are never chlorinated. Their chemical shifts are able to integrate the effects of multiple substitutions at the 2,2',3,3',4,4', etc. positions. For all three structural backbone types, these nonchlorinated carbons provide a global, quantum-mechanically-determined response to the degree and position of chlorinated substitution elsewhere in the molecule, a better indicator than would be provided by any one of the chlorinated carbon positions directly associated with the toxic effect. Therefore, selection of statistically significant bins at 135, 138, and 142 ppm supports the proposition that the QSDAR TEF model is built on a scientifically valid foundation.
The 10 bins selected for the QSDAR 50% REP2005 model are also listed in Supplementary Data, "Discussion of Model Results." They were not identical to those used in the 11-bin TEF model, though several appeared in both models. For the 10-bin REP model, the same kinds of structure–activity rationale could be used to reach the same conclusion: there was a demonstrable correlation between the spectral features selected and the structure–activity relationships upon which the validity of modeling depended.
This fact supports the work presented here but also suggests a caveat: one should be cautious in using the model for predicting the biological activity of drastically different compounds, especially those kinds for which the model was provided no inferential basis during training. For example, using the 11-bin TEF or the 10-bin REP model one would probably not obtain an accurate estimate of toxicity for polybrominated dioxins or furans. To address this question we have defined conservatively our estimate of the applicability domain for these models.
Inconsistencies between the TEF and REP predictions were of greater magnitude and more common for the unknown furans than for the dioxins and PCBs. This might be due to the fact that furans occupy a region of structural diversity between dioxins, and PCBs. Bins associated with furans are sometimes also occupied by signals from dioxins or PCBs. The models developed in this project are examples of 1D-QSDAR. The patterns contain no explicit structural information, so confusion of meaning can be associated with occupancy in some cases. When other factors are equal, 3D-QSDAR models, which incorporate structural information as interatomic distances associated with occupied NMR bins, are able to improve model performance for PCBs, furans, and dioxins (Beger et al., 2002). 3D-QSDAR might not improve on the r2 and q2 values shown here, but could improve external validation quality, especially for furans.
The results for predicting biological activity of tetrachlorophenothiazine pressed the limits for prediction beyond the structural variability range of the model training set and beyond our estimated applicability domain. Qualitatively, the large predicted REP and the very small predicted TEF could make sense when compared with the rationale behind tetrachlorophenothiazine's proposed use as a therapeutic drug (Fried et al., 2007): very high receptor affinity (modeled by REPs) and facile biodegradation (implicit in the training values of the TEF model). However, examination of predicted TEF and REP2005 values for other nonzero TEF congeners (see "Unknown TEF and REP Predictions" in Supplementary Data) did not show a consistent tendency for REPs to exceed TEFs, so that the relative differences in the predicted values for tetrachlorophenothiazine may represent chance rather than an interpretable effect. As should be expected, the difference between predicted and assayed values was considerably greater for the tetrachlorophenothiazine than for those typically found in the model training set congeners and for the two dioxins that were CALUX-assayed in this work.
The two congeners’ REP assessed by both CALUX and QSDAR is greater than that of about half the 29 congeners that have been assigned nonzero TEFs. Based on the measurements and models presented here, it would be hard to argue without in vivo testing that these two congeners should be assigned zero weight when conducting dioxin risk assessments.
Considering that the TEF scale covers five orders of magnitude, we contend that even if errors were equal to or greater than one order of magnitude, QSDAR-predicted TEFs or REPs would be useful to screen and, after CALUX confirmation, to nominate candidates for in vivo toxicology tests. An overestimate of toxicity, even by a factor of 10, would not be a problem because neither QSDAR predictions nor CALUX measurements would be used for regulatory decisions or risk assessment without in vivo validation and official approval. However, underestimates by factors of 1000 or more might have public health significance. This may actually be the case for some of the congeners presently defined as having zero TEFs. We propose the QSDAR methodology for rapidly, objectively, and easily estimating TEFs and REPs for all congeners of chlorinated dioxin-like compounds. These results argue for a reassessment of the list of congeners deemed "dioxin-like."
1,3,7,8-TCDD is an example of a compound that possesses serious toxic potency but is currently assigned a zero TEF. That it has been assigned a zero TEF might be surprising because of the reports of its significant binding to AhR in rats (Mason and Safe, 1986) and guinea pigs (Holcomb et al., 1998) as well as its known metabolic conversion to 2-hydroxy- and 3-hydroxy-tetrachloro-p-dioxins (Lans et al., 1993; Petroske et al., 1997). Such hydroxylated metabolites are competitors with the plasma thyroid hormone transport protein transthyretin in vivo in rat and in vitro in human systems (Lans et al., 1993) and they are reported to induce physiological changes (body and organ weights) and toxicological responses (aryl hydrocarbon hydroxylase induction) in rats (Mason and Safe, 1986). We acknowledge that significant metabolite toxicity is not equivalent to dioxin-like character and does not itself justify the inclusion of this congener in the TEF scheme. We suggest the possible inclusion of 1,3,7,8-TCDD in the TEF system not because of its toxic metabolites, but for two other reasons in addition to the predictions and assays presented in this work:
- (1) Because of its demonstrated presence in some food samples (we provide in Supplementary Data two examples of food analyses in which this TCDD and several others occur.)
- (2) Because there is some evidence that–for low level exposures–differences in persistence among dioxin and furan congeners may not be the same as those found in typical occupational or accident cohort exposures. If one is continuously exposed to zero-TEF congeners at low levels, then even rapid metabolism may not eliminate destructive dioxin-like effects compared with the persistent congeners. (We discuss this issue in Supplementary Data, "Evidence Supporting a Possible Reconsideration of TEF-based Risk Assessment.")
- (2) Because there is some evidence that–for low level exposures–differences in persistence among dioxin and furan congeners may not be the same as those found in typical occupational or accident cohort exposures. If one is continuously exposed to zero-TEF congeners at low levels, then even rapid metabolism may not eliminate destructive dioxin-like effects compared with the persistent congeners. (We discuss this issue in Supplementary Data, "Evidence Supporting a Possible Reconsideration of TEF-based Risk Assessment.")
In our study, a number of expensive, time-consuming steps were performed before two CALUX REPs could be obtained. First, the compounds had to be synthesized and purified, which was technically demanding and time consuming. After being quantitatively dissolved, each compound's purity was determined by GC/MS analysis followed by calculations. The postsynthesis steps were neither trivial nor cheap because of: (1) variations in the mass spectrometry protocol that had to be applied to determine the type and quantity of impurities that are normally ignored in regulatory dioxin analysis; (2) the expensive instrumentation; and (3) the well-trained and experienced operators of both the CALUX system and the mass spectrometer. Finally, although CALUX did not involve use of laboratory animals, it took several days to complete the experimental procedure and perform the REP calculations. However, the time and expense involved with CALUX were much less than those that would be required for animal experimentation.
In contrast, every aspect of the QSDAR method is computational, obviating needs for synthesis, identity confirmation, purity determination, dilution and NMR analysis of congeners. The virtual process allows quick spectral prediction for large numbers of congeners. The MLR pattern recognition model-building calculations are facile and rapid and do not require the use of a sophisticated computer platform. Once a rugged model has been developed and validated, it can generate a reliable toxicity estimate for any unknown within the applicability domain and do so the moment its predicted NMR spectrum is submitted to the model.
|
CONCLUSIONS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
Patterns in NMR spectra can be effectively used to model biological activity. Hundreds of dioxin-like chemicals, having been produced in massive quantities for decades, are now ubiquitous in the environment and the food chain. Because very few of them have been isolated, purified, and individually tested for biological activity, the potential for QSDAR as a practical screening and prioritizing tool seems clear. CALUX, a reliable and relatively simple in vitro assay, can be used for a second level quantitative confirmation. Two independently determined results suggesting the same conclusion can provide an objective basis for directing limited resources, to select candidates for in vivo toxicological assessment. In silico and in vitro nomination protocols may enable practical risk assessment for toxic end points in which members of a large chemical family exhibit different degrees of toxicity operating through a common mechanism.
|
SUPPLEMENTARY DATA |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
Supplementary data are available online at http://toxsci.oxfordjournals.org/.
Available on the web is a glossary of acronyms used in this paper. We are providing as separate documents a more detailed explanation of the CALUX-assay mechanism and a discussion of features and limitations for some non-QSDAR modeling methods.
Regarding methods and results, we are web-publishing an Excel spreadsheet, "Binned spectra 29 TEFs and two unknowns 020602007," that identifies (via an abbreviated code that omits the repeated terms "chloro" and "dibenzo") the congeners used in this study. It catalogs the binned NMR spectra of the various congeners, with populations shown in bold fonts for the eleven bins incorporated into the model TEF model and in italics for the ten used in the REP2005 model. (When a bin was used for both models it appears in bold italics.) This data was used to generated Figures 1 and 2. The identities of each congener having a symbol plotted on a figure appear in column A. By comparing a symbol's (x, y) coordinates to corresponding ordered pairs of negative logarithms (in columns C, BH for TEFs and BK, BM for REP2005s) one can deduce which congener the symbol represents.
We also provide three spreadsheets showing detailed results of cross-validation and external validation exercises. The spreadsheet file "TEF CV 10% Averages and STDs" reports cross-validation results for the ten models of the training set congeners depleted by three but with all 11 bins from the original model retained. The spreadsheet file "REP CV 10% Averages and STDs" reports cross-validation results for the ten models of the training set congeners depleted by three but with all 10 bins from the original model retained. The spreadsheet file "Log TEF CV 10% external validation tests" contains data for the ten Leave-Three-Out TEF models based on recalculating the set of optimal bins for each model.
Supplementary data includes a Table of Predictions for every congener in the models' Applicability Domain. Congeners that meet rather arbitrarily selected criteria for concern (discussed above) are shown in orange fonts.
Supplementary data, "Evidence Supporting a Possible Reconsideration of TEF-based Risk Assessment," also provides a more detailed discussion of results and two chromatograms showing a large number of non-2,3,7,8-TCDDs observed in seafood. The identifiable congeners include a peak in each chromatogram for the 1,2,3,7-TCDD modeled and assayed as an unknown in this paper. This data is part of a discussion of results shown by van den Berg in 1983. They show that congeners not presently regarded as dioxin-like because of their rapid detoxification and elimination in high dose experiments may not be so rapidly eliminated under low dose exposures. This supports the need to reconsider which congeners should be included in the TEF system used for risk assessment of chronic, low level exposures in food.
|
FUNDING |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
National Center for Toxicological Research; and the Arkansas Regional Laboratory, U.S. Food and Drug Administration.
|
NOTES |
|---|
Disclaimer: Views presented here do not necessarily reflect those of the U.S. Food and Drug Administration.
|
ACKNOWLEDGMENTS |
|---|
We greatly appreciate the gift of purified 1,3,7,8-TCDD and 1,2,3,4,7-PeCDD from Dr Steven Safe, Texas A & M University, without which the work could not have been done. We also acknowledge Dr Antony Williams and Brent Lefebvre for the beta-test use of ACD/Labs’ 13C predictor software.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONCONCLUSIONSSUPPLEMENTARY DATAFUNDINGREFERENCES |
|---|
Beger RD, Buzatu DA, Wilkes JG. Combining NMR spectral and structure data to form models of polychlorinated dibenzodioxins, dibenzofurans, and biphenyls binding to the AhR. J. Comput.-Aided Mol. Design (2002) 16:727–740.[CrossRef]
Beger RD, Wilkes JG. Models of polychlorinated dibenzodioxins, dibenzofurans, and biphenyls binding affinity to the aryl hydrocarbon receptor using 13C NMR data. Chem. Inf. Comput. Sci. (2001) 41:1322–1329.[CrossRef]
Behnisch PA, Hosoe K, Sakai S-i. Brominated dioxin-like compounds: In vitro assessment in comparison to classical dioxin-like compounds and other polyaromatic compounds. Environ. Int. (2003) 29:861–877.[CrossRef][Web of Science][Medline]
Birnbaum LS, DeVito MJ. Use of toxic equivalency factors for risk assessment for dioxins and related compounds. Toxicology (1995) 105:391–401.[CrossRef][Web of Science][Medline]
Brown DJ, Chu M, Overmeire IV, Chu A, Clark GC. Determination of REP values for the CALUX bioassay and comparison to the WHO TEF values. Organohal. Compd. (2001) 53:211–214.
Buzatu DA, Beger RD, Wilkes JG, Lay JO Jr. Predicting toxic equivalence factors from 13C nuclear magnetic resonance spectra for dioxins, furans, and polychlorinated biphenyls using linear and non-linear pattern recognition methods. Environ. Toxicol. Chem. (2004) 23:24–31.[CrossRef][Web of Science][Medline]
Covaci A, Koppen G, van Cleuvenberg R, Schepens P, Winneke G, van Larebeke N, Nelen V, Vlietinck R, Schoeters G. Persistent organochlorine pollutants in human serum of 50–65 years old women in the Flanders Environmental and Health Study (FLEHS). Part 2: Correlations among PCBs, PCDD/PCDFs and the use of predictive markers. Chemosphere (2002) 48:827–832.[Medline]
Fried KW, Schneider CM, Schramm K-W, Datta A, Chahbane N, Corsten C, Powell DR, Lenoir D, Kettrup A, Terranova P, et al. From dioxin to drug lead—The development of 2,3,7,8-tetrachlorophenothiazine. ChemMedChem (2007) [Published Online: 30 Mar 2007].
Hamers T, van Schaardenburg MD, Felzel EC, Murk AJ, Koeman JH. The application of reporter gene assays for the determination of the toxic potency of diffuse air pollution. Sci. Total Environ. (2002) 262:159–174.[CrossRef]
Hawes LC, Su SH, Harris M, DeVito MJ, Walker NJ, Farland WH, Finley B, Birnbaum LS. Development of a refined databasee of mammalian relative potency estimates for dioxin-like compounds. Toxicol. Sci. (2006) 89:4–30.
Holcomb M, Yao C, Safe S. Biologic and toxic effects of polychlorinated dibenzo-p-dioxin and dibenzofurans congeners in the guinea pig. Biochem. Pharmacol. (1998) 37:1535–1539.[CrossRef]
Hooper K, Hayward D, Chu M, Anderson M, Farland W, Lucier G, Clark G. CALUX results correlate with GC/MS data from Kazakstan breast milk samples and support new TEF values. Organohal. Compd. (2000) 45:236–239.
Koppen G, Covaci A, van Cleuvenbergen R, Schepens P, Winneke G, Nelen V, Schoeters G. Comparison of CALUX-TEQ values with PCB and PCDD/F measurements in human serum of the Flanders Environmental and Health Study (FLEHS). Toxicol. Lett. (2001) 123:59–67.[CrossRef][Web of Science][Medline]
Lans MC, Klasson-Wehler E, Willemsen M, Meussen E, Safe S, Brouwer A. Structure-dependent, competitive interaction of hydroxyl-polychlorobiphenyls, -dibenzo-p-dioxins, and -dibenzofurans with human transthyretin. Chem. Biol. Interact. (1993) 88:7–21.[CrossRef][Web of Science][Medline]
Mason G, Farrell K, Keys B, Piskorska-Pliszczynska J, Safe L, Safe S. Polychlorinated dibenzo-p-dioxins: Quantitative in vitro and in vivo structure-activity relationships. Toxicology (1986) 41:21–31.[CrossRef][Web of Science][Medline]
Mason G, Safe S. Synthesis, biologic and toxic effects of the major 2,3,7,8-Tetrachloro-p-dioxin metabolites in the rat. Toxicology (1986) 41:153–159.[CrossRef][Web of Science][Medline]
Murk AJ, Legler J, Denison MS, Giesy JP, van de Guchte C, Brouwer A. Chemical-activated luciferase gene expression (CALUX): A novel in vitro bioassay for the Ah receptor active compounds in sediments and pore water. Fundam. Appl. Toxicol. (1996) 33:149–160.[CrossRef][Web of Science][Medline]
Petroske E, Huwe JK, Feil VJ, Larsen GL. Identification of NIH-shifted metabolites of 1,3,7,8-tetrachloro-p-dixin in the rat by NMR comparison with synthesized isomers. Chemosphere (1997) 34:1549–1555.[Medline]
Safe S. Polychlorinated biphenyls (PCBs): Environmental impact, biochemical and toxic responses, and implications for risk assessment. Crit. Rev. Toxicol. (1994) 24:87–149.[Web of Science][Medline]
Safe S. Limitations of the toxic equivalency factor approach for risk assessment of TCDD and related compounds. Teratogen. Carcinogen. Mutagen. (1997/1998) 17:285–304.[CrossRef]
Safe SH. Development, validation, and problems with the toxic equivalency factor approach for risk assessment of dioxins and related compounds. J. Anim. Sci. (1998) 76:134–141.
Safe S. Polychlorinated biphenyls (PCBs), dibenzo-p-dioxins (PCDDs), and dibenzofurans (PCDFs) and related compounds: Environmental and mechanistic considerations which support the development of toxic equivalency factors (TEFs). Crit. Rev. Toxicol. (2002) 21:51–88.[CrossRef]
Stahl BU, Kettrup A, Rozman K. Comparative toxicity of four chlorinated dibenzo-p-dioxins (CDDs) and their mixtures. I. Acute toxicity and toxic equivalency factors. Arch. Toxicol. (1992) 66:471–477.[CrossRef][Web of Science][Medline]
Sugawara Y, Gee SL, Sanborn JR. Development of a highly sensitive enzyme-linked immunosorbent assay based on polyclonal antibodies for the detection of polychlorinated dibenzo-p-dioxins. Anal. Chem. (1998) 70:1092–1099.[Medline]
Tsutsumi T, Amakura Y, Nakamura M, Brown DJ, Clark GC, Sasaki K, Toyoda M, Maitani T. Validation of the CALUX bioassay for the screening of PCDD/Fs and dioxin-like compounds in retail fish. Analyst (2003) 128:486–492.[CrossRef][Medline]
van den Berg M, Birnbaum L, Bosveld ATC, Brunström B, Cook P, Feeley M, Giesy JP, Hanberg A, Hasegawa R, Kennedy SW, et al. Toxic equivalency factors (TEFs) for PCBs, PCDDs, PCDFs for humans and wildlife. Environ. Health Perspect. (1998) 106:775–792.[Web of Science][Medline]
van den Berg M, Birnbaum LS, Denison M, De Vito M, Farland W, Feeley M, Fiedler H, Hakansson H, Hanberg A, Haws L, et al. The 2005 World Health Organization reevaluation of human and mammalian toxic equivalency factors for dioxins and dioxin-like compounds. Toxicol. Sci. (2006) 93:223–241.
van den Berg M, Olie K, Hutzinger O. Chemosphere (1983) 12:537–544.
van Overmeire I, Chu M, Brown D, Clark G, Carbonnelle S, Goeyens L. Application of the CALUX bioassay for the determination of low TEQ values in milk samples. Organohal. Compd. (2001) 50:64–66.
van Overmeire I, Goeyens L, Beernaert H, Srebrnik S, De Poorter G, Baeyens W, Clark G, Chu M, Chu A, Chu D, et al. A comparative study of GC-HRMS and CALUX TEQ determinations in food samples by the Belgian Federal Ministries of Public Health and Agriculture. Organohal. Compd. (2000) 45:196–199.
Windal I, Van Wouwe N, Eppe G, Xhrouet C, Debacker V, Baeyens W, De Pauw E, Goeyens L. Validation and interpretation of CALUX as a tool for the estimation of dioxin-like activity in marine biological matrixes. Environ. Sci. Toxicol. (2005) 39:1741–1748.
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||