Environ. Sci. Technol. 2001, 35, 2282-2287
A Molecular Topology Approach to Predicting Pesticide Pollution of Groundwater FRED WORRALL† Department of Geological Sciences, University of Durham, Science Laboratories, South Road, Durham, DH1 3LE, UK
Various models have proposed methods for the discrimination of polluting and nonpolluting compounds on the basis of simple parameters, typically adsorption and degradation constants. However, such attempts are prone to site variability and measurement error to the extent that compounds cannot be reliably classified nor the chemistry of pollution extrapolated from them. Using observations of pesticide occurrence in U.S. groundwater it is possible to show that polluting from nonpolluting compounds can be distinguished purely on the basis of molecular topology. Topological parameters can be derived without measurement error or site-specific variability. A logistic regression model has been developed which explains 97% of the variation in the data, with 86% of the variation being explained by the rule that a compound will be found in groundwater if 6χvP < 0.55. Where 6χp is the sixth-order molecular path connectivity. One group of compounds cannot be classified by this rule and prediction requires reference to higher order connectivity parameters. The use of molecular approaches for understanding pollution at the molecular level and their application to agrochemical development and risk assessment is discussed.
Introduction The occurrence of pesticides in aquifers is well recorded throughout the world and has become one of the major diffuse pollution risks to the world’s drinking water supplies. With this pollution has come legislation to limit the maximum allowable concentrations of pesticide compounds in water for human consumption (1). With rising public concern regarding pesticide pollution and the instigation of environmental legislation the environmental fate and behavior of pesticides has become the concern of the agrochemicals industry with environmental fate and behavior criteria now crucial for pesticide registration (2). Consideration of the environmental behavior has become an integral part of the agrochemical development process and can influence the choice between compounds of equal pesticidal activity. This requirement has facilitated the development of a range of approaches for prescreening those compounds that represent an unacceptably high risk of pollution from those that are acceptable. The simplest of these methods rely on anecdote and experience; for example, ref 3 proposes parallel classifications of compounds based on adsorption (Koc (4)) and soil half-life (DT50) values. This approach produces broad bands of classification with arbitrary boundary values where small changes in the adsorption or degradation parameters † Corresponding author phone: ++44 (0)191 374 2535; fax: ++44 (0)191 374 2510; e-mail:
[email protected].
2282
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 35, NO. 11, 2001
may cause large changes in classification, which are well within the measured variability of these parameters. The majority of prescreening methods rely on deriving scores and indices. Perhaps the earliest of these methods (5) uses the solubility of the compound of interest to generate a mobility score. Later methods tend to use more than one parameter (6). A range of other techniques have been applied (7, 8). By far the most used assessment method is that of Gustafson (9) which is based on Californian data (10). This work shows that for the pesticides known to be used in the study area it is possible to discriminate between those compounds that leach to groundwater and those that do not purely on the basis of the adsorption coefficient (Koc) and the soil half-life (DT50) of the compound. From this observation, a groundwater ubiquity score is derivedsthe GUS score. This approach is attractive because it is derived solely from actual observations of the process of interest, i.e., groundwater pollution, rather than from assumptions regarding the important physical processes or from anecdotal evidence. However, the interpretation of the score remains informal and the value of one score relative to another lacks any particular meaning. Worrall et al. (11) treated the same Californian groundwater as a binary outcome (i.e. compound either leaches to groundwater or not) and in combination with several possible explanatory variables (adsorption, degradation, hydrolysis and solubility) derived the probability that a compound permeates to groundwater in concentrations above the limit of detection. This approach enables meaningful interpretation of compounds relative to each other. All these approaches however assume that the values of the parameters used are both without measurement error and are constant in the environment. An examination of pesticide property databases (USDA pesticide properties database (12) and (13)) shows that variability in the two most commonly used parameterssKoc and DT50scan be as much as several orders of magnitude. Indeed, Wooff et al. (14) showed that the Gustafson approach (9) could not be directly transferred to UK data. However, taking account of the variability, combinations of parameters could be chosen such that most compounds could be classed as either a leacher or a nonleacher. This suggests that what had been previously assumed by many authors to be constants dependent only on the nature of the compound is in fact highly site-specific. If the behavior of compounds is to be understood purely in chemical terms then allowance for this site-specific variability must be made. Without detailed knowledge of the application the above screening models could be misleading, potentially open to abuse, and could obscure chemical and molecular interpretations. However, Wooff et al. (14) developed a new statistical methodology to discriminate between leaching and nonleaching compounds on the basis of adsorption and degradation parameters, suggesting that despite the parameter variability, compounds and their environmental behavior might still be distinguishable based upon their chemical properties. Worrall et al. (15) showed that using these methods in combination, adsorption and degradation data gave highly significant results and showed that the derived results held true for UK groundwater. The result shows that it should be possible to distinguish between those compounds that leach and those that do not purely on the basis of chemical parameters. In turn, this means not only that accurate prescreening models can be developed but also that pollution 10.1021/es001593g CCC: $20.00
2001 American Chemical Society Published on Web 04/25/2001
of groundwater by pesticides may be understood in terms of molecular properties. This paper sets out to achieve both.
Approach and Methodology A typical approach used for relating chemical behavior to molecular properties is the use of quantitative structure activity relationships (QSAR). QSAR approaches have been extensively used in organic chemistry to understand reaction mechanisms and by the pharmaceutical sciences to understand drug action and facilitate drug development. Within the environmental sciences there has been less extensive use of this approach, but they have been used to predict the following: soil sorption coefficients (16); association coefficients with dissolved humic substances (17); the Henry’s law constant (18); bioconcentration factors (19); biodegradation rates (20); and acute toxicity (21). Many approaches are taken, some use multivariate statistical techniques, especially multiple linear regression (22), but increasing use has been made of neural networks (23) and expert systems (24). The models derived take a variety of approaches to what they predict, in some cases the QSAR is used to classify compounds (23), and develop a neural network to distinguish between weakly and highly degradable compounds. Other models identify structural elements that can be associated with, for example, a compound being biodegradable (25). Methods using multiple linear regression predict values of parameters, but as Langenberg et al. (26) suggest many published QSARs have inconvenient end points, e.g, classification into fast and slow degrading compounds rather than prediction of actual degradation rates, that can be of little use if QSARs are to be considered in risk assessment. Equally, Langenberg et al. (26) suggest that many QSARs are limited by the narrowness and number of the compounds for which they are derived. Just as the approach to deriving QSAR models varies so too does the range of descriptors used in them. Early approaches use the octanol-water partition coefficient (Kow) (17), but this parameter is empirical and can itself be predicted by a QSAR approach. Quantum molecular parameters have been used to predict the mineralization rate of s-triazines (27) and Koc (28). Structural fragments can be used both to develop a model, e.g. for biodegradation (29), or in combination with other molecular parameters, e.g. for adsorption (30). By far the most popular molecular parameter to be used is molecular connectivity (e.g. for biodegration (31), and for soil adsorption (32)). Quantum, fragment, and connectivity approaches are not prone to site-variability or measurement error. Such approaches to understanding either degradation or adsorption are of limited use in understanding pollutant transport and occurrence, but only indirectly, as they only predict a parameter that may or may not be important in predicting environmental fate of a compound. It is far more efficient to predict the property of interest directly, i.e., the occurrence in groundwater. One simple approach would be to combine QSAR descriptions of adsorption and degradation in the manner suggested by previous results (11). However, since it is known that adsorption and degradation parameters are highly variable within the environment, any combination of QSARs based on such parameters will be prone to the same errors. The more direct approach is to discriminate between leachers and nonleachers on the basis of molecular parameters. A set of connectivity parameters can be derived for compounds which have been included in groundwater surveys. The compounds that leach can then be distinguished from those that do not leach on the basis of these molecular properties by means of logistic regression. Logistic regression is the most appropriate technique for predicting a binary
outcome (leacher vs nonleacher) from continuous explanatory variables (e.g. molecular connectivity). This method transforms from a probability scale (0, 1) to the scale of continuous variables (∞, -∞). The transformation used is the logit transform, y ) log(θ/(1 - θ)) where θ ) the probability of reaching groundwater in measurable quantities (taken as 0.1 µg/L corresponding to the EC drinking water standard (1)). This probability should not be interpreted as a probable frequency of detection but rather the probability of finding the compound at least once. The transformed parameter y can then be linearly related to the chosen explanatory variables. This regression method does not use a least-squares fitting method but rather uses maximum likelihood estimation. Linear discriminate analysis may be used to classify a binary response to one of two groups given a set of explanatory variables. However, discriminant analyses does not offer a probability scale and predictions may fall outside acceptable regions. In the same way, neural networks do not offer a probabilistic scale of classification and thus the facility to interpret results relative to each other is lost. In addition logistic regression allows the significance of parameters included in the model to be assessed, thus aiding the mechanistic interpretation of the model. This study chose to work with molecular connectivities. Molecular connectivity was first derived by Randic (33) and later a more generic version was developed by Kier and Hall (34). Molecular connectivity is defined as h
χ)
∑
v )-1/2 (δiv * δjv ‚‚‚ δh+1
χ ) chi,
δ ) delta
where δv ) the vertex degree from the adjacency matrix (valence corrected), i, j ... refer to pairs of adjacent atoms, and the reciprocal square root products are summed across all possible subgraphs of the hydrogen suppressed graph of the molecule. This means that connectivity can be calculated for paths, for clusters, or for path-clusters depending upon how the graph of a molecule is subdivided. For the purpose of this study connectivity parameters are expressed as hχvm, where h is calculated from zero to ninth order and where m is designated p, pc, or c depending upon whether it is calculated for path fragments (p), path-cluster fragments (pc), or cluster fragments (c). The v refers to the fact that all connectivity parameters are valence corrected. For each compound in the study, 31 connectivity parameters were calculated. For further details of the development and calculation procedures for molecular connectivity see ref 35. Models were developed using data from the groundwater monitoring program of ref 36. This program monitored 56 compounds in 303 boreholes across 12 states in the midwest of the United States during 1991-1992. The compounds in the survey (Table 1) cover a wide range of properties, solubilities vary from 0.003 to 280 000 mg/L in water, and given the caveat that these values can vary by an order of a magnitude, the Koc varies from 1.5 to 160 000 L/kg and the aerobic soil half-life from 1 to 3800 days. The survey includes 47 compounds for which the molecular connectivity could be calculated, small fumigant compounds, metabolites, and compounds containing nitro groups were excluded. The selected compounds cover the range of common pesticides and contain a broad range of functional groups (Table 1). The models were validated against a second data set Johnson (37) and (38), updating the work of Wilkerson et al. (10) upon which the results of both Gustafson (9) and Worrall et al. (11) were based. These are data that were not available to ref 36, and, again excluding fumigant compounds and those with nitro functional groups, these data give results for 41 compounds: 27 of these compounds are common to both of data sets, yielding 61 compounds altogether. VOL. 35, NO. 11, 2001 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
2283
TABLE 1. List Compounds Used in This Study with the Pollutant Classification as Given in Refs 38 and 36a common name alachlor aldicarb ametryn atrazine azinphos-methyl bentazone bromacil butylate carbaryl carbofuran chloramben chlordane chlorothalonil chlorpyrifos chlorthal cyanazine dacthal ddt diazinon dicamba dimethoate disulfoton diuron endosulfan EPTC ethoprop fenamiphos fonofos gamma-HCH heptachlor linuron
class of pesticide
(38)
(36)
common name
chloroacetanilide carbamate s-triazine s-triazine organophosphate
x x x x NR x x NR x x x x x x x x NR x x x x x x x NR x x x x x x
x NR x x x x NR x x x NR NR NR x NR x x NR x x x x NR NR x x NR x x NR x
malathion methiocarb metolachlor metribuzin molinate naled napropamide oxamyl pebulate permethrin phorate picloram prometon prometryn pronamide propachlor propanil propargite propazine silvex simazine tebuthiuron terbacil terbufos terbutryn thiobencarb triallate 2,4-D 2,4-DP 2,4,5-T
uracil thiocarbamate carbamate carbamate s-benzoic acid organochlorine organochlorine s-benzoic acid s-triazine pthalate organochlorine organophosphorus s-benzoic acid organophosphorus organophosphorus phenylurea organochlorine thiocarbamate organophosphorus organophosphorus organophosphorus organochlorine organochlorine phenylurea
class of pesticide organophosphate chloroacetanilide s-triazinone thiocarbamate organocphosphorus carbamate thiocarbamate organochlorine organophosphorus pyridinecarboxylic acid s-triazine s-triazine chloroacetanilide organochlorine s-triazine organochlorine s-triazine phenylurea organochlorine organophosphate s-triazines thiocarbamte thiocarbamate organochlorine organochlorine organochlorine
(38)
(36)
x x
x NR x x x NR x NR x x x x x x x x x x x x x x x x x x x x x x
x NR x NR x NR NR x x x x NR x NR NR NR x x NR NR NR NR NR NR x NR NR
a Common name and class of compound refer to the Pesticide Manual (13) where IUPAC names and CAS numbers can be found. NR ) not reported in that study; x ) reported as a nonleacher; and x ) reported as a leacher.
Results The best-fit logistic regression model was calculated in a stepwise manner so that only additional parameters that gave statistically significant improvements to the model were included. The model was also tested for its stability and classification rate by exchange of parameters. Parameters excluded in the stepwise regression procedure were exchanged for parameters previously excluded as a means of ensuring the fit of the chosen model. Fitting the model to the 47 compounds in ref 36, for which connectivities were computed, yields the following estimated model
log(θ/(1 - θ)) ) 2.969 0χvp - 30.37 3χvp + 82.62 4χvp (0.08) (0.03) (0.03) 75.52 5χvp + 16.22 1χvpc - 56.13 2χvpc + 65.48 3χvpc (0.04) (0.08) (0.04) (0.06) 44.34 4χvpc - 283.4 7χpc (1) (0.06) (0.05) where θ ) the probability of leaching to groundwater in measurable quantities, and hχvm is as defined above. The values in the brackets are the P values, i.e., probability of the coefficient being zero. The constant term is found not be significant from zero (at the 95% level), and its removal did not affect the classification. This model gives 97% concordance with the data and 90% correct classification. Correctness of classification is judged against the 50% probability of being a leaching compound and the high concordance suggests that a number of those compounds misclassified by this relationship are only just misclassified. The data from the Californian data set ((37, 38) Table 1) show a high degree of correspondence with the Midwest 2284
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 35, NO. 11, 2001
data set. Of the 27 compounds common between the two data sets only three of them differ in their leaching status between the two. By applying the binomial theorem, assuming that the probability of being a leaching or nonleaching compound is random, it gives the probability of the observed pattern occurring by random is 0.0002%. This represents strong additional confirmation of the finding of refs 17 and 18 that leaching and nonleaching compounds have distinct chemical properties that could be discerned above the influence of site variability. Examination of data from refs 37 and 38 shows that applying eq 1 to only the data from California gives 91% concordance with the new data. Combining the datasets from the Midwest and California and then finding the best-fit model gives the following result:
log(θ/(1 - θ)) ) 0.53 χvp - 2.25 χvp + 9.67 4χvp (0.11) (0.04) (0.01) 14.03 χvp + 1.72 χvpc - 7.46 χvpc + 14.27 χvpc (0.04) (0.39) (0.08) (0.05) 11.18 4χvpc - 62.18 7χpc (2) (0.03) (0.04) This model gives 82% correct classification on all the data. It fits slightly better for nonleachers, than leachers, but the difference is smalls84% correct classification for nonleachers, as opposed to 79% for leachers. The P-values show that the term in χvpc is not very significant, but it is included because it does give an improvement in the classification rate and for comparison with eq 1. To simplify this relationship the best possible bivariate combination of all the explanatory variables was chosen by comparing the classification of pairwise combinations of all calculated variables for the combination of both the California
v FIGURE 1. 6χpv vs 7χpc in comparison to observation of leaching behavior for Midwest (36) and California (38) datasets (compounds are divided between the datasets in which they occur). Isoprobability lines are given that the compound will not leach to groundwater in measurable quantities.
TABLE 2. List Compounds Used in This Study with Their Topological Descriptors as Used in Eq 3 common name
6χ
alachlor aldicarb ametryn atrazine azinphos-methyl bentazone bromacil butylate carbaryl carbofuran chloramben chlordane chlorothalonil chlorpyrifos chlorthal cyanazine dacthal DDT diazinon dicamba dimethoate
0.7478 0.1175 0.5519 0.389 1.6819 1.0661 0.4821 0.6144 0.476 0.7834 0.2287 4.2335 0.4344 1.0097 0.4311 0.4115 0.593 1.0221 0.9868 0.2449 1.0237
p
7χ
pc
0.0076 0 0.0114 0.0026 0.1673 0.0465 0 0 0.003 0.0541 0 0.3878 0 0.042 0 0.0054 0 0.021 0.0723 0.0008 0
common name
6χ
disulfoton diuron endosulfan EPTC ethoprop fenamiphos fonofos gamma-HCH heptachlor linuron malathion methiocarb metolachlor metribuzin molinate naled napropamide oxamyl pebulate permethrin
3.331 0.3595 3.4652 0.6462 2.8748 1.4141 1.8204 0.9988 3.0537 0.3653 1.8416 0.5573 0.7562 0.4712 1.0719 1.0293 0.6068 0.2646 0.6683 1.0074
and Midwest datasets:
log(θ/(1 - θ)) ) 2.19 - 3.38 6χvp - 30.77 χvpc (0.01) (0.02) (0.21)
(3)
This equation gives 75% correct classification, with the fit being slightly better for leachers than nonleacherss79% correct for leachers and 73% for nonleachers. The advantage is that this relationship can be visualized (Figure 1, Table 2). This relationship better defines leaching compounds than the nonleachers. The fact that 6χvp appears in eq 3 and not eqs 1 and 2 shows that there is a large correspondence between molecular connectivities of similar orders. Taking
p
7χ
pc
0 0.0168 0.303 0 0 0.111 0 0 0.2271 0.013 0 0.0043 0.0123 0 0.0198 0 0.021 0.0017 0 0.0704
common name
6χ
phorate picloram prometon prometryn pronamide propachlor propanil propargite propazine silvex simazine tebuthiuron terbacil terbufos terbutryn thiobencarb triallate 2,4-D 2,4-DP 2,4,5-T
3.5683 0.288 0.3656 0.6032 0.4446 0.2894 1.2789 0.3657 0.4469 0.4453 0.4469 0.7387 0.3919 5.7013 0.5937 0.84 0.6163 0.355 0.4066 0.36
7χ pc
p
0 0 0.0101 0.0218 0.0527 0 0.0986 0.0048 0.0042 0.0069 0.0042 0.0084 0 0 0.0296 0.0619 0 0.0013 0.0051 0.0042
the 50% probability line as best discriminator between leachers and nonleachers a simple inequality can be calculated as a condition for a compound to leach to groundwater. A compound will leach to groundwater if h v χp
- 9.1 7χvpc < 0.86
(4)
This inequality correctly classifies 78% of the compounds. It is also easy to see that of these two parameters 6χvp is the most important, reflected in the low significance (high P value) level of the 7χvpc coefficient. A good separation could be achieved on the basis of this parameter alone. The logistic VOL. 35, NO. 11, 2001 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
2285
v FIGURE 2. 6χpv vs 7χpc in comparison to observation of leaching behavior for Midwest (36) and California (38) datasets (compounds are divided between the datasets in which they occur). The compounds highlighted refer to discussion in the text.
regression was calculated (eq 5):
log(θ/(1 - θ)) ) 2.38 - 4.3 χvp (0.04) (0.002)
(5)
This gives 74% correct classification with nonleachers being more correctly classified than leachers. Again if the 50% probability is considered the best discriminator of a leacher from nonleacher, then a single parameter condition for a compound to leach to groundwater can be given (Figure 2, line A): 6 v χp
< 0.55
(6)
This simple inequality correctly classifies 74% of the data. Viewing the combined data sets suggests that natural groupings exist in the data (Figure 2). The leaching compounds that were misclassified by eq 3 include diuron; linuron is also misclassified by this relationship, but it is a nonleaching compound. Comparing their connectivities (Table 2) shows that they are closely related structures, but these compounds are correctly classified by eq 1. That although eq 1 is best summarized by 6χvp this cannot encompass the complex 3Dbranching necessary to describe their behavior. Within the group of nonleaching compounds two trends can be observed (Trend A and B, Figure 2). The difference between these two trends is in their path-cluster connectivity. This difference is essentially between compounds that contain heterocycles or substituted aromatic rings (Trend A) and aliphatic compounds (Trend B). There is no evidence here to suggest that aromatic structures are inherently more or less leachable than aliphatic compounds.
Discussion These models and this type of approach represent a very constrained form of a screening model for the development of new agrochemicals, more constrained than has been traditionally been employed. Its simplicity is that no more information than is present in structural formula is required 2286
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 35, NO. 11, 2001
as input criteria. The model predicts the phenomena of interest, i.e., the probability of leaching, rather than parameters (e.g. adsorption) that would then have to be used in other models in order to predict the probability of leaching. The model results do not suffer from the variability and the measurement error of methods based purely on parameters such as Koc or soil half-life or indeed those QSAR approaches that predict such parameters. The model is based on robust information, i.e., purely on presence-absence data. This is an advantage in that it assumes little about the nature of the environmental data, but in this case it can be susceptible to individual pollution incidents, e.g. spills, that are unconnected to the natural properties of the site where the observation is made, the normal agricultural use of the chemical, or the properties of the compound observed. The effect of incidents of this type, such as accidental spills, can only be minimized by the examination of the largest number of compounds possible where false positive results can then be distinguished by analysis of the residuals from the statistical modeling. Future developments must look at the prediction of frequency of detection of compounds, e.g. how often in a region is a compound detected and not just whether it is detected at least once. However, the use of presence/ absence data is more in tune with the present regulatory framework than the prediction of frequency of detection. The present regulatory framework is focused on water quality standards such as the 0.1 µg/L standard (1). Such standards are based on analytical limits of detection rather then toxicity and as such are themselves a presence/absence indictor rather than a measure of actual concentration or frequency of detection. This may explain why that at no stage during the development of these models has the loading of the chemicals in the environment had to be considered. If actual concentration or frequency of detection were to be modeled rather than just the current presence/absence data, then loading in the environment may well become important. The above models also represent an advance in understanding the chemistry of pollution. Previous attempts to understand environmental fate in terms of the adsorption and degradation may be misleading, prone as they are to site
specific variability. An explanation of pollution on the basis of chemical parameters alone can only be made once a satisfactory explanation is found to describe the variabilities of the environmental data (14, 15). Lohninger (32) correlates zero- and first-order connectivity with both molecular volume and molecular weight which in turn directly relate to the solubility of the compound. Higher-order connectivities are more related to the scale of branching within the molecule. An increased degree of branching within a molecule will restrict microbial degradation by inhibiting access of enzymes to readily degradable functional groups and linkages and may cause steric hindrance of chemical degradation. The model derived in this study without direct reference to adsorption and degradation parameters shows that, to a first approximation, the effect of these two parameters can be combined in an understanding of molecular branching. For the majority of compounds a measure of the extent of branching appears sufficient (i.e. 6χvp). Even if many of the calculated molecular properties may be collinear it is clear that for more complex structures, e.g. ametryn, associations with path connectivities are not sufficient and that account must be taken of the path-cluster components of the structure. Once it has been proved possible to distinguish between polluting and nonpolluting compounds purely on the basis of chemistry, then several further possibilities present themselves, such as the use of quantum descriptors or fragment approaches. The possibility then arises that groundwater pollution by organic chemicals could be described in the same manner as Hammett equations (39). An example can be seen when comparing linuron and diuron. Eq 2 correctly predicts their behavior (their behavior is not well predicted by eq 3, Figure 2), but it is difficult to interpret differences in their connectivity in chemical terms, yet the difference is only a methyl group. A similar difference of a single methyl group between nonleaching and leaching compounds can be seen between 2,4-D and 2,4-DP; and silvex and 2,4,5-T, in each of these cases the compounds are correctly classified by eq 2 although not by eq 3. Such approaches will be a subject of future modeling. Molecular methods represent a useful forward in the prescreening of new compounds, risk assessment of compounds, and the potential for designing more environmentally friendly agrocemicals.
Acknowledgments The author would like to thank Dana Kolpin, U.S. Geological Survey, and Bruce Johnson, California Department of Pesticide Registration, for access to data collected and collated by their departments.
Note Added After ASAP Posting The ASAP version of this article was released on 4/26/01 before final corrections were made. Symbols in Table 1, footnote a were switched. This version has now been corrected and was posted on the web 5/8/01.
Literature Cited (1) EC Directive 80/778. (2) EC Directive 91/414. (3) Hollis, J. In Pesticides in soil and water; Walker, A., Ed.; British Crop Protection Council: Farnham, UK, 1991; Monograph no. 47, pp 165-174. (4) Helling, C. S.; Turner, B. C. Science 1968, 162, 562-563.
(5) Helling, C. S.; Kearney, P. C.; Alexander, M. Adv. Agronomy 1971, 28, 147-240. (6) Vereijken, P.; Wijnands, F.; Stol, W. Technical Report, AIR 3-CT920755, AB-DLO; Wageningen: Netherlands, 1995. (7) Meeks, Y. J.; Dean, J. D. J. Water Resources Planning Management 1990, 116, 693-707. (8) Halfon, E.; Galassi, S.; Bruggemann, R.; Provini, A. Chemosphere 1996, 33, 1543-1562. (9) Gustafson, D. I. Environ Toxicol. Chem. 1989, 8, 339-357. (10) Wilkerson, M. R.; Kim, K. D. The pesticide contamination prevention act: setting specific numerical values; California Department of Food and Agriculture: Sacremento, CA, 1986. (11) Worrall, F.; Wooff, D. A.; Seheult, A. H.; Coolen, F. P. A. Pesticide Sci. 1998, 54, 99-112. (12) Hornsby, A. G.; Wauchope, R. D.; Herner, A. E. Pesticide properties in the environment; Springer-Verlag: New York, 1996. (13) Tomlin, C. The pesticide manual, 10th ed.; British Crop Protection Council: Farnahm, UK, 1994. (14) Wooff, D. A.; Seheult, A. H.; Coolen, F. P. A.; Worrall, F. In Statistics for the Environment 4; Barnett, Stein, Feridun-Turkman, Eds.; Wiley: Chichester, 1999; pp 337-354. (15) Worrall, F.; Wooff, D. A.; Seheult, A. H.; Coolen, F. P. A. J. Geol Soc. 2000, 157, 877-884. (16) Sabljic, A. Environ Sci. Technol. 1987, 21, 358-366. (17) Sabljic, A.; Gusten, H.; Verhaar, H.; Hermens, J. Chemosphere 1995, 31, 4489-4514. (18) Nirmalakhandan, N. N.; Speece, R. E. Environ. Sci. Technol. 1988, 22, 1349-1357. (19) Govers, H.; Ruepert, C.; Aiking, H. Chemosphere 1984, 13, 227236. (20) Boethling, R. S.; Gregg, B.; Frederick, R.; Gabel, N. W.; Campbell, S. E.; Sabljic, A. Ecotox. Environ. Safety 1989, 18, 252-267. (21) Kier, L. B.; Hall, L. H. Bull. Environ. Contam. Toxicol. 1982, 29, 121-126. (22) Boethling, R. S. Environ. Toxicol. Chem. 1986, 5, 797-806. (23) Cambon, B.; Devillers, J. Quant. Struct.-Act. Relat. 1993, 12, 49-56. (24) Klopman, G.; Zhang, Z.; Balthasar, D. M.; Rosenkranz, H. S. Environ Toxicol. Chem. 1995, 14, 395-403. (25) Niemi, G. J.; Veith, G. D.; Regal, R. R.; Vaishna, D. D. Environ. Toxicol Chem. 1987, 6, 515-527. (26) Langenberg, J. H.; Peijnenburg, W. J. G. M.; Rorije, E. SAR QSAR Environ. Res. 1996, 5, 1-16. (27) Kaune, A.; Sixt, S.; Bruggemann, R.; Altschuh, J.; Kettrup, A. In Ecoinforma; Alef et al., Ed.; 1997; Vol. 12, pp 476-480. (28) Reddy, K. N.; Locke, M. A. Weed Sci. 1994, 42, 453-461. (29) Howard, P. H.; Boethling, R. S.; Stiteler, W. M.; Meylan, W. M.; Hueber, A. E.; Beaumant, J. A.; Larosche, M. E. Environ. Toxicol. Chem. 1992, 11, 593-603. (30) Meylan, W.; Howard, P. H.; Boethling, R. S. Environ. Sci. Technol. 1992, 26, 1560-1567. (31) Boethling, R. S.; Sabljic, A. Environ. Sci. Technol. 1989, 23, 672679. (32) Lohninger, H. Chemosphere 1994, 29, 1611-1626. (33) Randic, M. J. Am. Chem. Soc. 1975, 97, 6609-6615. (34) Kier, L. B.; Hall, L. H. Molecular connectivity in structure-activity analysis; Research Studies Press: Chichester, UK, 1986. (35) Silipo, C.; Vittoria, A. In Comprehensive Medicinal Chemistry, Quantitative Drug Design; Hansch, Sammes, Taylor, Eds.; Pergammon Press: NewYork, 1990; Vol. 4, pp 153-204. (36) Kolpin, D. W.; Goolsby, D. A.; Thurman, E. M. J. Environ. Qual. 1997, 24, 1125-1132. (37) Johnson, B. Setting revised specific numerical values - November 1988; California Department of Food and Agriculture: Sacremento, CA, 1988. (38) Johnson, B. Setting revised specific numerical values - April 1991; California Department of Food and Agriculture: Sacremento, CA, 1991. (39) Sykes, B. A guidebook to mechanism in organic chemistry, 6th ed.; Longman: Harlow, England, 1986.
Received for review August 15, 2000. Revised manuscript received March 6, 2001. Accepted March 9, 2001. ES001593G
VOL. 35, NO. 11, 2001 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
2287
