Estimation of Chemical Toxicity to
Wildlife Species Using Interspecies
S . R A I M O N D O , *, †P . M I N E A U ,‡A N D
M . G . B A R R O N†
U.S. Environmental Protection Agency, National Health and
Environmental Effects Laboratory, Gulf Ecology Division,
1 Sabine Island Drive, Gulf Breeze, Florida 32561, and
National Wildlife Research Centre, Science and Technology
Branch, Environment Canada, 1125 Colonel By Drive,
Ottawa K1S 5B6, Canada
Ecological risks to wildlife are typically assessed using
toxicity data for relatively few species and with limited
understanding of differences in species sensitivity to
contaminants. Empirical interspecies correlation models
were derived from LD50 values for 49 wildlife species and
quail (Coturnix japonica) and mallard (Anas platyrhynchos)
were determined to be good surrogates for many species
within the database. Cross-validation of all models
predicted toxicity values within 5-fold and 10-fold of the
actual values with 85 and 95% certainty, respectively. Model
robustness was not consistently improved by developing
correlation models within modes of action (MOA); however,
improved models for neurotoxicants, carbamates, and
direct acting organophosphorous acetylcholenesterase
inhibiting compounds indicate that toxicity estimates may
improve if MOA-specific models are built with robust
distance and cross-validation prediction success (?2)
297, df ) 12, p < 0.0001), with uncertainty increasing with
larger taxonomic distance between the surrogate and
predicted species. Interspecies toxicity correlations provide
a tool for estimating contaminant sensitivity with known
levels of uncertainty for a diversity of wildlife species.
Extrapolation between species and toxicity endpoints has
remained one of the key areas of uncertainty in both
ecological and human health risk assessment. For wildlife,
has typically been based on relatively limited toxicity data
and models have been used to extrapolate toxicity data
between wildlife species, including uncertainty or safety
factors, the derivation of species sensitivity distributions (1)
with or without body size scaling (2), and interspecies
correlation estimation (ICE) models (3). However, most of
the strategies involving the application of safety factors can
uncertainty in risk assessments (4). Strategies involving
species sensitivity distributions typically require well-
populated datasets and devolve, when data are limiting, to
either a safety factor approach that is empirically based (1)
or the use of pooled variances established for a large group
in species sensitivity distributions vary with both scientific
and regulatory applications.
that describe the relationship between the acute toxicity
(LD50; mg/kg bodyweight) of a range of chemicals tested in
two species. If toxicity data are available for the surrogate
species, the toxicity to the predicted taxon can be estimated
using the specific ICE model for that species-taxon pair. ICE
models have been employed in toxicity extrapolation of
aquatic invertebrates and fish (6), but their use in wildlife
risk assessment has not been widely accepted. Earlier
development of ICE models did not show good correlations
used in model development. Additionally, acceptance of
interspecies regression models has been limited due to lack
of model validation studies (7). While sublethal endpoints
maybe more ecologically relevant, the development of ICE
models for chronic toxicity is problematic because of
substantially fewer toxicity values compared to LD50 data-
bases for both aquatic and wildlife species.
This study developed a comprehensive set of ICE models
and identified key sources of uncertainty in wildlife toxicity
a comprehensive database compiled from open literature
(8-12) and by governmental agencies of the United States
(U.S. EPA Office of Pesticide Programs) and Canada (Envi-
ronment Canada) (1, 13). The suite of models were cross-
validated and analyzed to determined if (i) ICE models with
significant regressions can be used to predict acute toxicity,
(ii) there are certain chemical mode of action (MOA) that
should not be included in ICE models, (iii) ICE models
improve when developed with data of specific chemical
distance of the predicted and surrogate species, and (v) the
standard test species, mallard, northern bobwhite (Colinus
virginianus), and Japanese quail, are adequate surrogates
for predicting toxicity to other wildlife species. Results of
of ICE models in wildlife risk assessment.
Materials and Methods
single oral dose LD50 values for 156 species and 951
chemicals. The data were collected from the open literature
of the United States (U.S. EPA; Office of Pesticide Programs
have toxicity values for three or more chemicals or (2)
chemicals were not present in two or more species. Of the
2454 toxicity records used in used in the final analyses,
Data were subjected to rigorous quality assurance guide-
lines and standardized by using only data for adult animals
and data for chemicals of technical grade or formulations
>100 mg/kg or <100 mg/kg) and duplicate records among
* Corresponding author phone: (850) 934-2424; fax (850) 934-
2402; e-mail: email@example.com.
†U.S. Environmental Protection Agency.
multiple sources were not included in model development.
When data were reported as a range (ie. 100-200 kg/mg; ref
8) or data were collected from multiple sources for a species
and chemical, the geometric mean of the values was used.
In cases where the range of minimum and maximum values
records for that chemical were removed for that species due
to their high variability.
Model Development and Validation. Models were de-
veloped using Model II least-squares regression in which
(3). An algorithm was written in S-plus (14) to pair every
or more common chemicals per pair were required for
taxa) ) a + b × Log10(surrogate species), where a and b are
that had a significant linear relationship (p-value < 0.05)
were used for further analyses, because models that are not
significant at the p < 0.05 level have an elevated probability
of performing a Type I error.
The uncertainty of each model was assessed using leave-
one-out cross-validation. In this method, each pair of LD50
removed from the original model. The remaining data were
used to rebuild a model and estimate the toxicity value of
species toxicity value. This method could only be used for
models with degrees of freedom equal or greater than 2 (N
g 4). To maintain uniformity among the large number and
diversity of models, the N-fold difference among each
estimated and actual value (nontransformed data) was
calculated and used to determine the fit of the estimated
toxicity value. The wildlife database was used to estimate
interlaboratory variation for wildlife species from the range
of multiple toxicity values for a single species and chemical.
The average range of multiple toxicity measurements for a
in the analysis and 6.4 ( 15.8 (N ) 286) when only data from
different sources were compared. Thus, a 5-fold difference
was deemed a good fit in the validation analysis. Each
removed data point was assigned to a prediction category
The categories were 5-fold (e5-fold), 10-fold (>5-fold, e10-
fold), 50-fold (>10-fold, e50-fold), and greater than 50-fold.
The cross-validation success rate was calculated for each
model as the proportion of removed data points that were
predicted within 5-fold of the actual value from models that
were statistically significant. In cases where the removal of
an xy data pair resulted in the development of a model that
were not included in the cross-validation success rate and
a p-value between 0.01 and 0.05 in the original model.
in the selection of robust models for toxicity estimation in
related to cross-validation success rate. Model MSE is the
unbiased sum of squared deviations of the prediction line
A simple linear regression was conducted to compare cross-
validation success rate with MSE for all models in which
cross-validation success rate could be calculated (N ) 484).
Model R2is the relative prediction power of the model and
describes the proportion of variation in the data explained
mode of action (MOA) was explored in two analyses. The
first analysis was conducted to determine if any chemical
MOA consistently failed to be predicted by ICE models. The
using only data from one MOA. For both analyses, each
chemical in the database was assigned to a broad MOA (ie.
acetylcholinesterase (AChE) inhibitors) and specific MOA
(e.g., carbamate mediated AChE inhibition). Mode of action
structure, including major moieties and functional groups,
the mechanism of acute toxicity, therapeutic category, and
pesticidal activity through a review of reports and peer
reviewed articles (e.g., 16-21), and Internet sources (e.g.,
12 broad MOA and 26 specific MOA categories (Supporting
Information 1). Organophosphorous acetylcholinesterase
phosphoramidates) were also subdivided into direct acting
(containing an oxon moiety) and indirect acting chemicals
requiring metabolic activation to the oxon.
point predicted in the cross-validation analysis. For every
model, each removed y data point was associated with a
prediction success rate (5-fold, 10-fold, etc.) and chemical
pooled (total N ) 11846). For every broad and specific MOA
category and OAI subgroup, a binomial test was conducted
of the actual value for g50% of the time it was used in a
models could not be used to predict toxicity for each MOA.
The test statistic was the number of observations of an MOA
within the 5-fold category of prediction, the sample size N
was the total number of times the MOA was predicted in all
models and the probability of failure (p) was 0.5, corre-
sponding to the hypothesized rate of failure g50% of the
time it was used in a model (22).
The second chemical MOA analysis was conducted to
determine if models improved if built from only one MOA
or subgroup. For all models with df g50 (N ) 64, 32 species
pairs), individual models were built for the most abundant
and specific MOAs (carbamates and OAIs), and the OAI
subgroups. Models built with data from a specific MOA or
subgroup required a minimum of five data points to be
included in the analysis to avoid including models with
extremely small sample sizes that may not adequately
MSE of models comprised of all data were compared to
models built from data within an MOA or subgroup using a
MOA or subgroup compared to models developed with all
data were also explored qualitatively using the northern
bobwhite-mallard model. MOA-specific models were built
with northern bobwhite and mallard as the surrogate and
predicted species, respectively, for AChE inhibition, car-
bamates, OAIs, direct acting OAIs, and indirect acting OAIs.
Uncertainty of models developed with all data was
compared to taxonomic distance by assigning each model
predicted species. Models were rated such that surrogate
and predicted species within the same genus ) 1; family )
2; order ) 3; class ) 4; phylum ) 5 (i.e., birds vs mammals).
compared among taxonomic distance categories and
cross-validation category using a ?2test for differences in
There were 560 significant models (p < 0.05) developed for
49 species from the complete dataset (Table 1). The avian
surrogate species with the most models were red-winged
blackbird (Agelaius phoeniceus, 34), mallard (31), rock dove
(Columba livia, 30), Japanese quail (28), and ring-neck
pheasant (Phasianus colchicus, 28). The mammal species
with the most models were Norway rat (Rattus norvegicus,
26) and the mouse (Mus musculus, 19). Information on all
models, including statistics and parameters necessary for
toxicity estimation (slope, intercept, degrees of freedom,
number of chemicals in model, mean square error, sum of
squares) are provided in the Supporting Information 2.
Cross-validation was conducted for 536 models, 52 of
which resulted in the development of models that were no
longer significant at the p < 0.05. Thus, 484 models were
successfully cross-validated. Cross-validation of all models
within 5-fold of the actual value for 85% of all removed data
points and within 10-fold for 95% of removed points (N )
11846). Model MSE was significantly related to cross-
- 73.5x (R2) 0.64, df ) 482, p < 0.0001; Figure 1a). Based
on this relationship, MSE of 0.22 and 0.15 corresponded to
cross-validation success rate of 85 and 90%, respectively.
Model R2was also related to cross-validation success rate
and is described by the relationship y ) 64.2 + 34.6x (R2)
0.27, df ) 482, p < 0.0001; Figure 1b). Based on this
validation success rate of 85 and 90%, respectively. Model
rate (R2) 0.64) than the relationship described in the R2
selecting between two models.
The first MOA analysis determined that there was not an
actual values more than 50% of all times the MOA was
predicted by the combined set of models. On average, data
of the actual values for 85% of all occurrences.
In the second MOA analysis, models built with data from
anesthesia, AChE inhibition, OAIs, and indirect acting OAIs
MOAs were not improved over models built with all data.
For these models, values of R2were significantly lower and
MSE was not significantly different from models built with
all data. Models built with neurotoxicants, carbamates, and
direct acting OAIs that were significant at the p ) 0.05 level
TABLE 1. Species Used in ICE Models with the Total Number of Data Points and Statistically Significant Models (p < 0.05) for
common name scientific name
models common namescientific name
red billed quelea
eastern screech owl
white-crowned sparrow Zonotrichia leucophrys
yellow headed blackbird Xanthocephalus
fulvous whistling duck
golden-crowned sparrow Zonotrichia atricapilla
red billed quelea
FIGURE 1. Relationship of cross-validation success to (a) model
mean square error (MSE) and (b) model R2.
were improved over models built with all data based on
significantly lower MSE (Table 2). For all MOA or subgroups
specific data resulted in the loss of a significant relationship
for some models. For example, based on the criteria that
models developed from a possible 64 models. Of those 58
carbamate models, 46 were significant at the p ) 0.05 level
These results are supported by the qualitative analysis of
the northern bobwhite/mallard model built for all data
and 1.7 times smaller than the model built with all data,
respectively; however, the carbamate model was not sig-
nificant at the p ) 0.05 level. The MSE for models built with
markedly different than the MSE for the model built with all
to the all data model. Additionally, the Northern bobwhite/
mallard model was not significant when built solely with
indirect acting OAIs (Figure 2).
There was a strong relationship between taxonomic
distance and cross-validation prediction category (?2) 297,
df ) 12, p < 0.0001). There was a decrease in the percentage
genus predicted within 5-fold of the actual value for 100%
of all data points, where models built for two species within
the same phylum (bird vs mammal) predicted within 5-fold
of the actual value for 76% of all data points. There was an
increase in the percentage of data points in the other
prediction categories (10-fold, 50-fold, >50-fold) with in-
creasing taxonomic distance (Table 3).
Based on the results of model uncertainty analyses,
surrogacy was evaluated for the avian and mammal species
for which the most models were developed (Table 4). Rock
than 20 models with MSE e 0.22 or R2g 0.6, indicating that
these three species have the potential to provide good
pheasant, and house sparrow (Passer domesticus) are po-
tentially good surrogates for more than 15 models based on
low MSE or high R2. Northern bobwhite, common grackle
quelea (Quelea quelea) provided good surrogacy for over 10
species each based on low MSE. For mammal species, both
the Norway rat and mouse had 10 models with MSE e 0.22
or R2g 0.6 (Table 4).
significant for 560 models built for 49 species. Model MSE
TABLE 2. Mean Square Error (MSE) and R2((std) Compared for Models Built with Data from All MOAs with Those Built with
Data from Only One MOA/chemical Class or Subgroup. Only Models That Were Significant (p < 0.05) When Built with One Mode
of Action/Chemical Class or Subgroup Were Included in the Analyses.
mode of action/chemical class models developed significant models model parameter all data models MOA specific modelsdft p-value
102 -2.24 0.03
-5.18 < 0.01
aAChE is acetylcholinesterase.bOrganophosphorous-type acetylcholinesterase inhibitors, including organophosphates, phosphorothioates,
phosphonates, phosphoroamidates, etc.cDirect inhibition of AChE.dIndirect acting requires metabolic transformation to oxon form for AChE
bobwhite and mallard.
MOA/chemical class-specific models for northern
TABLE 3. Number of Observations within Each Prediction
Category for Models Grouped by Closest Shared Taxona
cross-validation prediction category
observations 5-fold 10-fold 50-fold
aNumbers in parentheses are the percentage of observations in
each prediction category for each shared taxonomic level.
and R2are linearly related to model robustness and should
be used to guide risk assessors in selection of appropriate
cross-validation success rate than R2and should be used as
the primary criterion in model selection. High cross-
validation success rate (g90%) for models built for species
within the same genus, family, and order demonstrate that
ICE models are generally most robust when they possess
close taxonomic distance (within order), small model MSE
(approximately 0.22 or less), and large model R2value
(approximately 0.6 or greater), and high cross-validation
success rate (approximately 85% or greater).
Interspecies extrapolation models use surrogate species,
often standard test organisms, to estimate toxicity of a
chemical to wildlife species for which no toxicity data exists.
In North America, tests are usually conducted on mallards
and northern bobwhite, whereas European countries often
use Japanese quail as their standard test species (13). As a
surrogate species, the number of significant models was 31
for mallard, 20 for northern bobwhite, and 28 for Japanese
high R2, and/or high cross-validation success rate, demon-
strating their potential to be used as avian surrogates for
estimating toxicity to wildlife species with limited or no
toxicity data for a specific chemical. Several other species,
pheasant, were surrogate species for more than 20 models
and offer potential to be used as surrogates for toxicity
with increasing taxonomic distance, estimating toxicity to
mammals from surrogate bird species is not recommended.
models, respectively, for mammal species and are recom-
mended for estimating toxicity to mammals in wildlife risk
assessment. Fewer ICE models for mammal species is a
reflection of a bias toward birds within the dataset and does
not necessarily reflect poor surrogacy of mammal-based
ICE models can be updated to include new chemicals and
species in the suite of ICE wildlife models. Future research
test data become available.
Model uncertainty analyses using leave-one-out cross-
validation provides only an estimate of generalization error
and is not a true validation of model fitness. Particularly for
models built from small datasets, a small change in the data
can cause a large change in the model (23). Although
bootstrap validation or leave-v-out cross-validation may
Additionally, because the model fitness criterion (5-fold) is
tests rather than model prediction error, the leave-one-out
method provided a vehicle for comparing estimated toxicity
values among a diversity of different models.
attributed to chemical class/MOA-specific toxicity, which
can exceed 2 orders of magnitude (1). Reduced model MSE
and increased model R2were used to determine if MOA-
specific models were improved over models built from all
data. Based on these criteria, only models built from
neurotoxicants, carbamates, and direct acting OAIs were
improved over models built with all data. However, many
MOA-specific models, including those developed for MOAs
p ) 0.05 level (Table 2). Although increased model p-value
is concomitant with reduced degrees of freedom, it also
signifies a reduction in model robustness and confidence.
While there is evidence to suggest that models may be
is necessary to ensure model robustness is preserved.
Chemical class and MOA-specific ICE models may provide
significant improvement in sublethal toxicity prediction;
is problematic because of relatively limited datasets com-
pared to acute toxicity.
The comparison of MOA-specific models with models
built from all data was only conducted for a small group of
MOA and subgroups; however, more than half of the MOA-
specific models did not demonstrate marked improvement
over models built with all data. The ICE models show that
species generally have high degree of toxicity in another
species. Also, ICE models describe interspecies variation on
a logarithmic scale, in which a large degree of change in
actual toxicity values is necessary for a small change in the
linear model. This comparison might yield different results
if the compared models were of equal sample sizes (i.e., by
selecting a random subsample of all data that was equal in
TABLE 4. Summary of the Surrogate for Which the Most ICE Models Were Developed and the Number of Models of Each
Surrogate Fitting Each Model Parameter (i.e., MSE, R2, etc.) Criteriaa
MSE MSE e 0.22
success rate g 85%
distance e 3
ICE models for risk assessment. While it is recognized that
toxicity (25), it does not appear necessary to develop MOA-
specific models to achieve confident ICE model estimates
for use in risk assessment. The results of these analyses
suggest that ICE models built from all available data can be
Safety factors, the generation of species sensitivity dis-
compound extrapolation models based on body size alone
have been the most commonly used methods for extrapolat-
ing toxicity between wildlife species in ecological risk
while SSDs have been used to identify generic receptors of
putative sensitivity. Fairbrother (26) and others have recom-
mended greater reliance on empirical data and robust
statistical approaches because of the high variation in
organism responses and the sometimes arbitrary nature of
safety factors. Body size scaling in birds and mammals is an
empirical approach that extrapolates from an LD50 (mg/kg)
based on organism weight. Scaling factors can vary greatly
for different chemicals and where chemical-specific factors
are unknown, an average is applied to extrapolate toxicity
among wildlife species (1-2). ICE models provide an
MOA and taxonomic level. There may be grounds to
incorporate scaling in future ICE models, isolating the
proportion of the overall inter-species variance that is the
from phylogeny and other scale-independent factors. In
Unfortunately, such an approach will prove data-intensive
few species have been tested.
The application of ICE models requires selection of the
most appropriate surrogate species based on a review of
cross-validation success rate, and taxonomic distance. In a
hypothetical example, the acute toxicity of a chemical to the
red-winged blackbird needs to be estimated in light of
available data on the toxicity of the chemical to five species:
duck, or common grackle (see Supporting Information 2).
The grackle has the closest taxonomic distance (2), low MSE
and high degrees of freedom (53), and is the best potential
surrogate in this example. Japanese quail (MSE ) 0.19, R2)
distance ) 4) would be the next best surrogate based on low
MSE and high cross-validation success rate, followed by
northern bobwhite (MSE ) 0.22, R2) 0.63, df ) 45, cross-
validation success rate ) 87, taxonomic distance ) 4) and
mallard (MSE ) 0.27, R2) 0.48, df ) 80, cross-validation
whistling duck has the highest model R2(0.91) and lowest
MSE (0.05), low df (df ) 2) and high p-value (0.047) do not
make it as good of a surrogate as other species because the
range of applicability of the model is limited by the small
dataset and the significance of the relationship is weak.
In screening level assessments, high confidence in in-
terspecies toxicity estimation is necessary. To increase
be calculated for the expected toxicity value (15). Narrow
confidence limits represent high confidence that the model
species toxicity. ICE model have the most confidence, and
can only be cross-validated, within the range of data points
that were used to estimate the model. As such, using ICE
models to predict a toxicity value from surrogate species
and average values of the surrogate species used to derive
each model) and equations are provided in the Supporting
Information 2 to facilitate use of these models in risk
assessment. Additionally, the models developed here are
available as a predictive modeling tool (Web-ICE) through
the EPA Center for Exposure Assessment Modeling website
updates to these models will be provided as new data
becomes available. This Internet tool provides model in-
formation and user guidance, and calculates toxicity values
and confidence intervals for selected models. ICE models
can be used when toxicity data for a specific chemical are
available for a selected surrogate species, and there is an
existing model between the species pair of interest. Models
for application in wildlife risk assessment to reduce uncer-
tainty in determining species sensitivity. The application of
for extrapolating across species and for assessing relative
For assistance with database development, we thank Sonny
Mayer (U.S. EPA, retired), and Thomas Steeger and Brian
Baril and Brian Collins (Environment Canada). Rick Bennet
(U.S. EPA) provided a valuable critique of the manuscript.
Also, thanks to our support personnel: Deborah Vivian,
Marion Marchetto, Anthony DiGirolamo, Christel Chancy,
Gulf Ecology Division).
Supporting Information Available
Additional information on chemicals used in model devel-
described in this study are available in Supporting Informa-
Contam. Toxicol. 2001, 170, 13-74.
(2) Mineau, P.; Collins, B. T.; Baril, A. On the use of scaling factors
to improve interspecies extrapolation of acute toxicity in birds.
Reg. Toxicol. Pharmacol. 1996, 24, 24-29.
(3) Asfaw, A.; Ellersieck, M. R.; Mayer, F. L. Interspecies Correlation
Estimations (ICE) for Acute Toxicity to Aquatic Organisms and
Wildlife. II. User Manual and Software, EPA/600/R-03/106;
United States Environmental Protection Agency: Washington,
extrapolation in birds and mammals and a proposal for long-
term toxicity data. Ecotoxicology 2005, 14, 817-832.
and mammals. Environ. Toxicol. Chem. 1997, 16, 1785-88.
(6) Dyer, S. D.; Versteeg, D. J.; Belanger, S. E.; Chaney, J. G.; Mayer,
F. L. Interspecies correlation estimates predict protective
environmental concentrations. Environ Sci. Technol. 2006, 40,
(7) Colton, J. A. Toxicity-extrapolation models. In Ecological
Modeling in Risk Assessment: Chemical Effects on Populations,
Ecosystems, and Landscapes; Pastorok, R. A., Bartell, S. M., Download full-text
Ferson, S., Ginzbury, L. R., Eds.; Lewis Publishers: Boca Raton,
FL, 2002; p 181, Chapter 12.
of Pesticides to Wildlife; Resource Publ. 153; United States Fish
and Wildlife Service: Washington, DC, 1984. p 90.
(9) Shafer, E. W., Jr.; Bowles, W. A., Jr.; Hurlbut, J. The acute oral
toxicity, repellency and hazard potential of 998 chemicals to
one or more species of wild and domestic birds. Arch. Environ.
Contam Toxicol. 1983, 12, 355-382.
(10) Shafer, E. W., Jr.; Bowles, W. A., Jr. Acute oral toxicity and
repellency of 933 chemicals to house and deer mice. Arch.
Environ. Contam Toxicol. 1985, 14, 111-129.
(11) Shafer, E. W., Jr.; Bowles, W. A., Jr. Toxicity, Repellency or
Phototoxicity of 979 Chemicals to Birds, Mammals and Plants;
Research Report No. 04-01; United States Department of
Agriculture: Fort Collins, CO, 2004; p 118.
(12) Smith, G. J. Pesticide Use and Toxicology in Relation to Wildlife:
Organophosphorus and Carbamate Compounds; Resource
Publication 170; United States Department of the Interior:
Washington, DC, 1987; p 171.
(13) Baril, A.; Jobin, B.; Mineau, P.; Collins, B. T. A Consideration of
Inter-Species Variability in the Use of the Median Lethal Dose
(LD50) in Avian Risk Assessment; Technical Report No. 216;
Canada Wildlife Service: Ottawa, Canada, 1994; p 12 and
(14) Insightful. S-plus Programmer’s Guide; Insightful Corporation:
Seattle, WA, 2001.
(15) Dowdy, S.; Weardon, S. Statistics for Research; John Wiley and
Sons: New York, 1991.
Nomenclature. A Guide to IUPAC Recommendations; Blackwell
Science: Oxford, UK 1998.
(17) Barron, M. G.; Hansen, J.; Lipton, J. Association between
contaminant tissue residues and adverse effects in aquatic
organisms. Rev. Environ. Contamin. Toxicol. 2002, 173, 1-
(18) Barron, M. G. Environmental Contaminants Altering Behavior.
Sons: West Sussex, UK, 2002; pp 167-186, Chapter 3.
Rahway, NJ, 1989.
D. E.; Drummond, R. A. Predicting modes of toxic action from
chemical structure: acute toxicity in the fathead minnow
(21) Wishart, D. S.; Knox, C.; Guo, A. C.; Shrivastava, S.; Hassanali,
M.; Stothard, P.; Chang, Z.; Woolsey, J. DrugBank: a compre-
hensive resource for in silico drug discovery and exploration.
Nucleic Acids Res. 2006, 34, D668-D672.
(22) Conover, W. J. Practical Nonparametric Statistics, 3rd ed; John
Wiley and Sons.: New York, 1999.
(23) Breiman, L. Heuristics of instability and stabilization in model
selection. Ann. Stat.s 1996, 24, 2350-2383.
(24) Efron, B. Estimating the error rate of a prediction rule:
improvement on cross-validation. J Am. Stat. Assoc 1983, 78,
(25) Mineau, P.; Whiteside, M. Lethal risk to birds from insecticide
use in the United States: a spatial and temporal analysis.
Environ. Toxicol. Chem. 2006, 25, 1214-1222.
(26) Fairbrother,A.Ecotoxicology: Useofsafetyfactorsinecological
risk assessment. In Encyclopedia of Ecology; Jørgensen, S. E.
Ed.; Elsevier: Amsterdam, The Netherlands, 2007, in print.
Received for review February 12, 2007. Revised manuscript
received April 23, 2007. Accepted April 26, 2007.