7

Development and Practical Application of Petroleum and DispersantInterspecies Correlation Models for Aquatic SpeciesAdriana C. Bejarano*,† and Mace G. Barron‡

†Research Planning, Inc., 1121 Park Street, Columbia, South Carolina 29201, United States‡U.S. Environmental Protection Agency, Gulf Ecology Division, 1 Sabine Island Drive, Gulf Breeze, Florida 32561, United States

*S Supporting Information

ABSTRACT: Assessing the acute toxicity of oil has generally relied onexisting toxicological data for a relatively few standard test species, which haslimited the ability to estimate the impacts of spilled oil on aquaticcommunities. Interspecies correlation estimation (ICE) models weredeveloped for petroleum and dispersant products to facilitate the predictionof toxicity values to a broader range of species and to better understandtaxonomic differences in species sensitivity. ICE models are log linearregressions that can be used to estimate toxicity to a diversity of taxa based onthe known toxicity value for a surrogate tested species. ICE models have onlypreviously been developed for nonpetroleum chemicals. Petroleum anddispersant ICE models were statistically significant for 93 and 16 uniquesurrogate-predicted species pairs, respectively. These models had adjustedcoefficient of determinations (adj-R2), square errors (MSE) and positive sloperanging from 0.29 to 0.99, 0.0002 to 0.311, and 0.187 to 2.665, respectively. Based on model cross-validation, predicted toxicityvalues for most ICE models (>90%) were within 5-fold of the measured values, with no influence of taxonomic relatedness onprediction accuracy. A comparison between hazard concentrations (HC) derived from empirical and ICE-based speciessensitivity distributions (SSDs) showed that HC values were within the same order of magnitude of each other. These resultsshow that ICE-based SSDs provide a statistically valid approach to estimating toxicity to a range of petroleum and dispersantproducts with applicability to oil spill assessment.

■ INTRODUCTION

Assessments of the potential toxicological effects of physicallyor chemically dispersed oils and dispersants have commonlyrelied on relatively few toxicity tests for a limited number ofaquatic species, primarily standard test species. In most cases,the sensitivity of species of concern is unknown, makinginformed decisions challenging. One approach to addressinguncertainty in species sensitivity has been the development andapplication of interspecies correlation estimation models(ICE),1 which are log−linear regressions between the acutetoxicity of two species for many paired toxicity data. Thesemodels can generate acute toxicity estimates by extrapolatingknown toxicity values from a surrogate species to species withunknown toxicity values. While these models are available formany chemicals,2−5 these have not been developed forpetroleum or dispersant products. ICE models can be used togenerate species sensitivity distributions (SSDs),2,6,7 which arecumulative distributions of toxicity data (e.g., median lethal,LC50, and effects concentrations, EC50) that allow forcomparisons of the relative sensitivities of across species.8

SSDs have been used to establish protective levels to specificchemicals9 by deriving hazard concentrations (HC) assumed toprotect a wide number of species with varying sensitivities. ICE-based SSDs have been developed for a number of chemicals,2,6,7

and have been shown to produce HC values similar to thoseused to derive water quality criteria.6

Recent research has shown that existing data from stand-ardized toxicity tests can be used to develop oil-specificSSDs,10−13 but species diversity is extremely limited comparedto single compounds.2,4,5 The objectives of the current studywere to determine if petroleum and dispersant-specific ICEmodels could be developed and used to generate SSDs andHC5 values with acceptable uncertainty, and to determineassociations between taxonomic relatedness and speciessensitivity.4 Previous studies have documented variability intest results attributable to differences in test conditions andexposure media prepared from the same source oil and/ordispersants.14−17 Therefore, this study focused on two distinctcategories of studies (spiked and continuous exposure) thatreported measured aqueous exposure concentrations ofpetroleum products (primarily light and medium crude oils)and several oil spill dispersants (primarily Corexits). Themodels and approach presented here can be used to estimatetoxicity to a broad range of species and assemblages which

Received: October 25, 2013Revised: March 17, 2014Accepted: March 21, 2014Published: March 28, 2014

Article

pubs.acs.org/est

© 2014 American Chemical Society 4564 dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−4572

should facilitate the development of protective environmentalconcentrations and improved assessment of the potentialconsequences of oil spills and dispersant use.

■ EXPERIMENTAL SECTION

Petroleum and Dispersant Toxicity Data Sources. Theprimary data source used in the development of ICE modelscame from a recently developed toxicity database,18 whichcontains quantitative information on the acute toxicity of crudeoils and dispersants from publically available literature. Thisdatabase was developed following a quality assurance andquality control (QA/QC) plan similar to that of a relateddatabase.19 QA/QC procedures included an evaluation of eachoriginal data source, removal of duplicate information, and anevaluation of currently accepted scientific names.Two types of data sets were queried and used in the

development of ICE models: petroleum hydrocarbon anddispersant toxicity data. The petroleum hydrocarbon data setwas comprised of toxicity data (LC50 and EC50) for aquaticspecies (primarily fish and crustaceans) derived from aqueousexposures to physically dispersed (i.e., water accommodatedfractions; WAF) or chemically dispersed (i.e., enhanced wateraccommodated fractions; CEWAF) oil. These toxicity datawere generated using a variety of oils under various weatheringstages, but most data were from light (e.g., Chirag, Forties,South Louisiana, Venezuelan) and medium (e.g., Alaska NorthSlope, Forcados, Kuwait, Prudhoe Bay) crude oils (42 oils of 48total oils). Only studies that reported LC50 and EC50 toxicityvalues on the basis of measured concentrations of analytes inthe aqueous exposure media were included in the developmentof these models. While rigorous data selection for petroleumhydrocarbons focused on test results reported on the basis ofmeasured total hydrocarbon content (THC; including C6−C36 carbon chains; 72% of the entire data set), other reportedmetrics (total polycyclic aromatic hydrocarbons, TPAHs; sumof PAHs and alkyl homologue groups) were also included. Thisdata set did not include results from studies reporting effectsconcentrations for specific THC carbon chains or PAHanalytes. Petroleum ICE models were developed using over1500 paired data points for a total of 136 unique species pairs.The dispersant data set was comprised of toxicity data (LC50

and EC50) for aquatic species exposed to dispersants, almostexclusively Corexit 9500 and Corexit 9527 (36% and 33% ofthe entire data set respectively). Toxicity data for otherdispersants (Corexit 7500; Corexit 7664; Corexit 9552;Nokomis; SlickAWay) were also included. Only studiesreporting LC50 and EC50 toxicity values on the basis ofmeasured dispersant concentrations in the exposure media wereincluded in model development. Dispersant ICE models weredeveloped using 286 paired data points for a total of 38 uniquespecies pairs.The majority of data included here were derived from

standard toxicity tests, including 96 h (65% of the entire dataset) and constant static (76% of the entire data set) laboratoryexposures. To develop petroleum and dispersant ICE models,surrogate and predicted species were paired from the sameoriginal data source. Pairing of surrogate and predicted speciestoxicity data was done only when tests for both species wereperformed under the same exposure conditions (i.e., same oil ordispersant product, exposure regime, analytical methods inchemical characterization), but with data independentlycollected for each species. A complete list of data sources and

core data are provided in the Supporting Information (SI)material.

Model Development and Verification. Because of thenumber of steps involved in the development, selection,verification and application of ICE models, a diagram isprovided (SI Figure S1) to facilitate the understanding of theapproach presented here. As several statistical methods wereused, the readers are encouraged to refer to key statisticalreferences20,21 including those describing the development ofICE models.2,5 Linear regression models were developed foreach pair of species in both, the petroleum hydrocarbon anddispersant database, containing at least 4 data points. Theselinear models are described by Log Pi = β0 + β1 × Log (Si),where Pi is the acute toxicity of the predicted species, β0 and β1are the intercept and slope, respectively, and Si is the acutetoxicity of the surrogate species, where both the independentand dependent variables are random and independent of eachother.22 The slope represents change in the response of toxicityvalues of the predicted species per unit change in the toxicityvalues of the surrogate species, where a slope of approaching 1indicates a similar response between two species. Only ICEmodels (p-value < 0.05) that passed both the F-test for theoverall fit of the regression equation and the t test for the slopeparameter significantly different from 0, were included infurther analyses. Prior to model validation, statisticallysignificant models were evaluated for potential influential datavia regression diagnostics analyses.20 The reliability andpredictive power of statistically significant ICE models with atleast 2 degrees of freedom was assessed using a leave-one-outcross-validation technique.23,24 Briefly, each data set was splitinto K subsets, equal to the number of pairs within each dataset, and models fitted K times, each time leaving out one datasubset from the larger training data. For each reduced data set,toxicity values of the predicted species were calculated andresponses for the deleted subset predicted from the model.Model uncertainty of the estimated toxicity value was calculatedvia the cross-validation success rate (or 1-bias correctedmisclassification error or sum of the square difference betweenobserved and estimated values), which is an estimate ofgeneralized error. Taxonomic relatedness for each surrogateand predicted species was assigned a numeric distance value4

(from same genus = 1 to same domain = 7), and modelprediction accuracy assessed as a function of taxonomicdistance.

Practical Application of ICE Models. Regressionparameters of statistically significant ICE models were usedto estimate toxicity values of several predicted species, using theknown toxicity of petroleum hydrocarbons or dispersants forone or more surrogate species as the model input. In allinstances, input data for surrogate species were independent ofthe data used to develop ICE models.10,18 Surrogate speciesincluded Americamysis bahia (mysid shrimp), Holmesimysiscostata (kelp forest mysid), Menidia beryllina (inland silverside)and Atherinops af f inis (topsmelt). ICE-based SSDs weregenerated for each category of petroleum or dispersant productusing the ICE estimated toxicity values, and in all cases theminimum acceptable number of species on an SSD was set tofive. The petroleum product categories included light crude oils(>31.1°API) (Forties, South Louisiana, Venezuelan crudes)and medium crudes (22.3−31.1°API) (Alaska North Slope,Arabian Medium, Kuwait and Prudhoe Bay crudes), while thedispersant categories included Corexit 7664, Corexit 9500, andCorexit 9527. Hazard concentrations (HC) assumed to be

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724565

protective of 99% or 95% of the species in the SSD (i.e., HC1and HC5, respectively) were computed from the modeled SSDfunction. All SSDs and HC values were derived using anapproach described elsewhere.25 Briefly, toxicity values werefitted to a log-normal distribution function, and SSDs randomlyresampled 2000 times to derive the HC1 and HC5 values andtheir associated 95% confidence intervals (95% CI). Only SSDsthat passed goodness of fit tests (α = 0.01) (the Anderson−Darling for SSDs with >7 species, and the Kolmogorov−Smirnov test statistics) were included in these analyses. SSDswere also generated using empirical data for the same categoriesof petroleum products and dispersants, as well as using acombination of empirical data and ICE-based toxicity values.SSDs with empirical data were generated by calculating thegeometric mean of all toxicity values available for each uniquespecies. The geometric was used because it gives less weight toinfluential data or potentially outliers, leading to conservativeestimates.8 To determine if ICE-based SSDs are a viablealternative to the estimation of SSDs where toxicity data arelimiting, ICE-based SSDs were compared to empirically basedSSDS using log-likelihood statistics.21 This approach comparesthe log-likelihood values of individual SSDs (e.g., M. beryllinaICE-based and empirical data) with the fitted SSDs of thecombined (pooled) model (e.g., M. beryllina ICE-based plusempirical data), testing the hypothesis (via the chi-squarestatistic) that these SSDs are derived from the same fitted log-normal curve. The same approach was used to compare ICE-based SSDs pairs. All the analyses above were performed usingthe R statistical platform (v. 2.13.2).22

■ RESULTS

Model Development and Verification. A total of 93petroleum ICE models for 29 surrogate species, and 16dispersant ICE models for seven surrogate species werestatistically significant (p-value <0.05 for both, the F-test forthe overall fit and the t test for the slope parameter) (SI TablesS1−S3, Figures S2 and S3). Models lacking significancetypically had limited paired data (22 petroleum and 5dispersant ICE models with <5 paired data points) limitingtheir use in further analyses.

The validity of statistically significant models was verifiedusing standard regression diagnostics procedures.20 Visualexamination of the standardized residuals of each significantmodel (Figure 1, left panel) showed that all residuals fell withina horizontal band centered on 0, indicating that the variance ofthe error term was constant. No obvious outliers were detected(±4 standardized residual values). Diagnostic analyses alsoindicated that the error term did not depart substantially from anormal distribution (Figure 1, right panel), as the coefficient ofdetermination (R2) between residuals and quantiles forpetroleum and dispersant ICE models were greater (0.997and 0.991, respectively) than those of the critical R2 value(0.985) (α = 0.05; n ≥ 100).26 A moderate departure fromnormality was noted at the end tails of the distributions,particularly for petroleum ICE models (14 data points) but wasconsidered minimal because removal of these data points didnot cause major changes in fitted values. In all cases, theabsolute change between fitted values with and withoutpotentially influential points ranged from 4 to 11% indicatingno disproportional influence on fitted values.Statistically significant models had an adjusted coefficient of

determination (adj-R2) ranging from 0.29 and 0.99, a MeanSquare Error (MSE; a measure of fit) ranging from 0.0002 to0.311, and a positive slope ranging from 0.187 to 2.665 (SITables S2 and S3). Examination of model parameter viabilityshows that 92% of all intercepts were between −0.75 and 0.75,and that 85% of all slopes were between 0.5 and 1.5. Thisrelatively narrow variability suggests similarities across mostmodels. Notable exceptions, where both the slope and interceptof the same model were outside these ranges included (1) thepetroleum ICE models for the surrogate-predicted pairsCalanus sinicus- Paracalanus aculeatus (both copepods species),Cyprinodon variegatus (sheepshead minnow)-A. bahia, andPontogammarus maeoticus (amphipod)-Calanipeda aquae dulcis(copepod); and (2) the dispersant ICE models for the pairs C.variegatus-A. bahia, A. af f inis-H. costata, and A. af f inis-Macrocystis pyrifera (giant kelp). In these deviating cases,except pairs involving A. af f inis, the predicted species wasgenerally more sensitivity than the surrogate species, but in allcases MSE were small (below 0.16).

Figure 1. Diagnostic analyses of homogeneity of variances (left) and normality (right) for petroleum and dispersant ICE models.

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724566

The reliability and predictive power of statistically significantmodels was further assessed by identifying the MSE cutoffassociated with a-priori cross-validation success rates of 85%and 95%,5 assumed to have moderate and high reliabilities,respectively (Figure 2). MSE cutoffs for petroleum and

dispersant models were 0.126 and 0.04, respectively. Overhalf of the petroleum ICE models (54 models) had MSEs≤0.04, while over 30% (32 models) had MSEs between 0.04and 0.126. By comparison, dispersant ICE models were equallydistributed among MSE cutoffs, suggesting that petroleum ICEmodels may be more robust.The high cross-validation success rate (a measure of

predictive power) was consistent with the agreement betweenpredicted and observed values for both ICE models types(Figure 3). For petroleum ICE models, 98% of the predictedvalues were within 5-fold difference of the observed data, withmost values (92%) being within 2-fold difference of theobserved data. There was no influence of taxonomic relatednesson prediction accuracy (p-value >0.05), and mean fold-difference values across taxonomic distances ranged from 1.00to 1.38 (SI Figure S4). Only 0.5% of predicted values (seven

observations) were >10-fold difference from the observedvalues. These outliers occurred only in theM. beryllina- A. bahia(taxonomic distance = 6), Daphnia magna (water flea)- Artemiasalina (brine shrimp) (taxonomic distance=4), and A. bahia- M.beryllina surrogate- predicted pairs (3, 3, and 1 observations,respectively). For dispersant ICE models, 99% of the predictedvalues were within 5-fold difference of the observed data, withmost values (88%) being within 2-fold of the observed data.There was no influence of taxonomic relatedness on predictionaccuracy (p-value >0.05), and mean fold-difference values were1.29 and 1.14 for taxonomic distances 6 and 7, respectively.The largest fold difference of 7 occurred in the A. bahia- M.beryllina surrogate- predicted pair.

Practical Application of ICE Models. Toxicity data forseveral petroleum products from an analysis independent of theresearch presented here10 were used to verify the applicabilityof ICE models. Two sets of SSD were developed using ICEmodels for A. bahia and M. beryllina as the surrogate species,with input surrogate concentrations from Barron et al.10 Thefirst set of SSDs used all ICE models regardless of theirreliability, while the second set used only ICE models withMSE <0.12. Statistical comparison of SSDs via (using the log-likelihood),21 showed that these two types of SSDs for severalpetroleum products and for both surrogate species, were notsignificantly different from each other (p-value >0.05). Thisindicates that exclusion of the least reliable models did notinfluence the shape of the SSD. No statistically significantdifferences were also found between A. bahia and M. beryllinaICE-based SSDs (p-value >0.05), or between either of thesetwo ICE-based SSDs and empirical SSDs10 (p-value >0.05;Figure 4). Estimated petroleum hydrocarbon HC5s from ICE-based SSDs were in general agreement with those derived fromSSDs using empirical data10 (SI Table S5), though values fromM. beryllina as the surrogate were generally larger (up to a 6fold larger) than those calculated from empirical data.A second verification approach utilized empirical toxicity data

for physically and chemically dispersed combined fromconstant and spiked exposures, as well as toxicity data forCorexit 9500 and Corexit 9727,18 which were used to generateSSDs (Figure 5). The selection of these data was driven by thefact that these SSDs shared data for two surrogate species: H.costata and A. bahia. SSDs were derived using all ICE models

Figure 2. Assessment of model reliability based on the relationshipbetween model mean square error (MSE) and cross-validation successrate for both, petroleum and dispersant ICE models.

Figure 3. Comparison of observed and ICE-predicted toxicity values for petroleum hydrocarbons and dispersants. The solid line represents the 1:1line (equal toxicity), while the dotted lines represent a 5-fold difference between these values. Bar chart on the right, display the fold-differencebetween observed and ICE-predicted toxicity values.

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724567

for these surrogate species regardless of model reliability.Because there were insufficient paired surrogate-predicted ICE

models for the same surrogate species to generate SSDs fordispersants (five species minimum), predicted ICE modelconcentrations were combined for H. costata and A. bahia.Curve comparison of empirically and ICE-based SSDs, usinglog-likelihood, showed no statistically significant differencesbetween A. bahia and H. costata ICE-based SSDs (p-value>0.05), or between either of these two ICE-based SSDs andempirical SSDs10,18 (p-value >0.05), with one exception: the H.costata ICE-based SSDs and the empirical SSDs18 for constantexposures to dispersed oil (p-value = 0.01; Figure 5). In mostcases, ICE-based models with H. costata and A. bahia assurrogates, produced smaller HC5s indicative of a slight modelbias toward overprotection of aquatic species. In all cases, HC5sfrom ICE-based SSDs were within the same order of magnitudeas HC5s from empirical SSDs. HC5s for Corexit 9500 andCorexit 9727 from ICE-based SSDs were also similar to thosereported by Barron et al.10

Hazard concentrations (96 h HC1 and HC5) were estimatedfor fresh petroleum and dispersant products for which data, onthe basis of measured concentrations, were available fromconstant static and spiked flow-through exposures (Table 1).These hazard concentrations were estimated using ICE-basedSSD with models from each of two surrogate species (A. bahiaand M. beryllina). An additional set of SSDs was alsoconstructed by combining all available empirical data plus

Figure 4. Comparison between species sensitivity distribution (SSD)curves for all petroleum products from Barron et al.10 (dashed line),and SSDs derived using all ICE models, regardless of their reliability,with A. bahia (black dots and lines) and M. beryllina (blue dots andlines) as the surrogate species. These curves are not significantlydifferent (p-value >0.05) from each other.

Figure 5. Comparison between physically and chemically dispersed oil (top), and dispersant (bottom) species sensitivity distribution (SSD) usingempirical18 and ICE-based models (red dots and lines; red dashed line from Barron et al.10 for dispersants only). SSDs using empirical data fordispersed oil were developed for constant and spiked exposures separately. SSDs were derived using all ICE models regardless of their reliability withA. bahia (black dots and lines) and H. costata (blue dots and lines) as surrogate species, combining models for these two species for dispersant SSDs(black lines). Within panels, SSDs are not significantly different (p-value > 0.05) from each other, except for the comparison of H. costata ICE-basedSSDs and the empirical SSDs for constant exposures to dispersed oil (p-value = 0.01; top left).

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724568

ICE toxicity values for one or more surrogate species (A. bahia,H. costata, M. beryllina and A. af f inis), keeping the smallestpredicted value for each unique predicted- species (SI FigureS5). Starting concentrations for surrogate species were those ofstudies reporting measured toxicity data,14,15,27−37 and excludedfrom model development. HC1s and HC5s for seven fresh oilsand by oil category ranged from 0.13 mg THC/L to 3.72 mgTHC/L, and from 0.16 to 4.92 mg THC/L, respectively.Venezuelan crude oil had the lowest HC values (more toxic),while Alaska North Slope oil had the highest values (less toxic).HC values from ICE-based SSDs with A. bahia as the surrogate

species were similar to those with M. beryllina, and weregenerally within the same order of magnitude of each other.HC values from SSDs that combined empirical data andpredicted toxicity values from one or more surrogate specieswere also similar to those derived with either A. bahia or M.beryllina. In all cases, spiked flow-through exposures producedHC values up to 12 times greater than HC values from staticexposures, with Venezuelan and Alaska North Slope crude oilhaving the largest and smallest differences, respectively,between these exposure types. SSDs using A. bahia or M.beryllina as surrogate species, and SSDs with empirical plus ICE

Table 1. Estimated HC1 and HC5 Petroleum Hydrocarbon (THC) Concentrations (mg/L) for Specific Petroleum andDispersant Products and Exposure Regimesa

petroleum/dispersantproduct

experimentalconditions

surrogate: A. bahiaHC1 (first row) | HC5 (second row)

(95% CI)

surrogate: M. beryllinaHC1 (first row) | HC5 (second row)

(95% CI)

empirical+ ICE-predictedHC1 (first row) | HC5 (second row)

(95% CI)

Alaska North constant static 0.55 (0.30−0.92) 3.03 (1.96−4.55) 0.45 (0.24−0.77)Slope 0.69 (0.44−1.02) 3.97 (2.93−5.35) 0.63 (0.37−0.99)

flow-through 1.03 (0.61−1.60) 3.72 (2.44−5.63) 1.27 (0.75−2.06)1.32 (0.93−1.86) 4.92 (3.69−6.57) 1.72 (1.09−2.58)

Arabian flow-through NA 0.59 (0.34−0.95) NA

Medium NA 0.71 (0.49−1.00) NA

Forties constant static 0.15 (0.07−0.28) 0.14 (0.06−0.29) 0.17 (0.07−0.34)0.18 (0.10−0.29) 0.16 (0.08−0.26) 0.18 (0.09−0.36)

flow-through 2.32 (1.41−3.66) 2.10 (1.36−3.16) 1.76 (1.05−2.72)3.06 (2.11−4.29) 2.70 (1.98−3.62) 2.28 (1.50−3.35)

Kuwait constant static 0.23 (0.11−0.41) 0.21 (0.10−0.37) 0.23 (0.12−0.41)0.28 (0.17−0.42) 0.23 (0.14−0.37) 0.26 (0.15−0.43)

flow-through 2.62 (1.58−4.29) 1.56 (1.01−2.31) 1.26 (0.77−1.9)3.49 (2.43−4.95) 1.97 (1.45−2.61) 1.57 (1.05−2.23)

Prudhoe Bay constant static NA 1.94 (1.24−2.84) NA

NA 2.48 (1.82−3.31) NA

flow-through 2.40 (1.40−3.90) 3.11 (2.01−4.68) 3.54 (2.24−5.41)3.19 (2.21−4.50) 4.09 (3.03−5.40) 4.52 (3.14−6.42)

South constant static 0.92 (0.51−1.49) 1.27 (0.80−1.94) 0.75 (0.43−1.20)Louisiana 1.15 (0.76−1.68) 1.58 (1.16−2.10) 0.97 (0.63−1.49)Venezuelan constant static 0.13 (0.06−0.24) 0.19 (0.09−0.36) 0.16 (0.07−0.31)

0.16 (0.09−0.25) 0.21 (0.12−0.35) 0.18 (0.09−0.32)flow-through 2.12 (1.27−3.38) 0.29 (0.14−0.50) 0.36 (0.19−0.61)

2.78 (1.94−3.93) 0.33 (0.20−0.50) 0.42 (0.25−0.66)Light crudesb constant static 0.19 (0.09−0.35) 0.44 (0.24−0.74) 0.22 (0.11−0.39)

0.24 (0.13−0.4) 0.53 (0.31−0.82) 0.29 (0.16−0.47)flow-through 0.84 (0.45−1.44) 0.64 (0.37−1) 0.36 (0.17−0.66)

1.07 (0.61−1.74) 0.78 (0.48−1.2) 0.47 (0.24−0.82)Medium constant static 0.62 (0.41−0.89) 1.21 (0.74−1.88) 0.27 (0.12−0.49)crudesc 0.71 (0.50−0.98) 1.50 (0.98−2.21) 0.35 (0.17−0.62)

flow-through 1.74 (1.04−2.85) 1.72 (1.09−2.63) 2.35 (1.52−3.54)2.25 (1.44−3.55) 2.20 (1.51−3.12) 2.98 (2.08−4.14)

Corexit 7664 flow-through NA NA 375 (182−744)789 (478−1145)

Corexit 9500 constant static NA NA 2.24 (1.25−3.73)3.33 (2.15−5.13)

flow-through NA NA 27 (16−42)40 (27−60)

Corexit 9527 constant static NA NA 1.09 (0.62−1.80)1.62 (1.02−2.49)

flow-through NA NA 5.51 (3.44−8.60)7.65 (5.27−11.02)

aHC Values were generated from ICE-based SSDs using surrogate-predicted ICE Models with at least five pairs per unique surrogate, as well as bycombining empirical and ICE predicted toxicity data from one or more surrogate species. Known surrogate concentrations include 96 h THCtoxicity data from physically dispersed fresh oil and fresh oil chemically dispersed with Corexit 9500, except for Kuwait Oil, which was chemicallydispersed with Corexit 9527. Surrogate toxicity data are the geometric mean of all measured values for the same species.14,15,27−37 NA, data notavailable. bForties + South Louisiana + Venezuelan. cAlaska North Slope + Arabian Medium + Kuwait + Prudhoe Bay.

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724569

toxicity values were not statistically different (p-value > 0.05)within petroleum products and by experimental condition.While there were insufficient dispersant ICE models for A.bahia or M. beryllina as surrogate species to produce ICE-basedSSDs, combined data from empirical and ICE toxicity valuesallowed the estimation HC values for three dispersants.

■ DISCUSSIONOne of the greatest challenges in evaluating potential impactsto aquatic communities during oil spills and impact assessmentcontinues to be uncertainties in species sensitivity. TheDeepwater Horizon oil spill further highlighted the need tounderstand the sensitivity of the diversity of aquatic species inthe deep ocean, pelagic and coastal areas with generally limitedexisting toxicity data.38 The current study provides bothpetroleum and dispersant-specific toxicity estimation modelsthat can be applied to a broad range of aquatic speciesassemblages using a SSD-based approach.The practical use of existing oil toxicity data has been limited

by the general lack of standardized laboratory practices,including differences in media preparation.13,14,39,40 Whilethese issues may have added uncertainty to the developmentof petroleum and dispersant ICE models, emphasis was placedon rigorously standardizing data selection with the sole intentof reducing the uncertainty introduced by differences inexperimental procedures across studies. As a result, and despiteinherent limitations with existing data, all statistically significantICE models had a relatively small MSE (range 0.0002−0.311),indicative of a robust model fit. As indicated elsewhere,6 andapplicable here, these robust ICE model relationships suggestthat the same mechanisms of action for each species pair maybe at play. Furthermore, MSE values associated with 85% and95% cross-validation success rates of ICE models were smaller(0.126 and 0.04, respectively) than MSE values associated withthe same cutoffs (0.22 and 0.15, respectively) for ICE modelswith wildlife species.5 In addition, predicted values from ICEmodels were generally within 5-fold difference of the observeddata, with most models being within 2-fold of the observeddata. These predicted-observed differences are well within thefold difference commonly found across laboratories (folddifference of 3)41 during optimal interlaboratory comparisonswith the same species. Taxonomic relatedness has beenpreviously shown to influence model fit and reliability of ICEmodels developed from chemicals with mixed modes ofaction.4,5 In contrast, petroleum and dispersant-specific ICEmodels showed no influence of taxonomic distance on modelaccuracy. These results suggest that ICE models developed withchemicals with a common nonspecific mode of chemical actionsuch as narcosis can be used to predict toxicity across a broadrange of taxa, and can help improve predictions over ICEmodels from with mixed modes of action.While SSDs have been used in the field of aquatic toxicology

and integrated into the regulatory framework,42 their use in oilspill research has been limited.10−12 SSDs and derivedbenchmarks can be used to protect untested species underthe assumption that their sensitivity is within the range ofsensitivities captured by the species in the SSD. Although SSDscannot replace toxicity testing, they can provide additionalinformation when the costs of toxicity testing are prohibitive orspecies-specific testing is restricted or not feasible (e.g.,endangered, rare, deepwater species). Furthermore, bothempiric and ICE-based SSDs can help inform resourcemanagers in their assessment of potential acute effects

associated with petroleum or dispersant products, particularlywhen data are limited. Moreover, data from concurrent toxicitytesting of rarely investigated taxa (e.g., corals, pelagic fish) and asurrogate species for which ICE models are available (e.g., A.bahia or M. beryllina), can be used to construct an ICE-basedSSD allowing for the placement in the curve of the species forwhich little toxicity data, facilitating comparisons of relativesensitivities. As demonstrated here, ICE models could be usedto augment estimates of benchmark concentrations from spikedflow-through exposures that may be more applicable to short-duration oil spills, and for which toxicity data are less availablein the scientific literature.13 As shown here and elsewhere,2,6,7

ICE-based SSDs can produce HC5 values similar to thosegenerated from empirical SSDs, adding reliability to the use ofICE models to augment toxicity data. Here, HC5s from ICE-based SSDs for petroleum and Corexit dispersants were within1 order of magnitude HC5s from SSDs with empirical data.10,18

While previous studies have recommended between 7 and 15species to develop reliable SSDs and associated benchmarks,2,7

the minimum number of species used here was 5. As a result,petroleum ICE-based SSDs could only be developed for 9surrogate species (including A. bahia and M. beryllina), or bycombining data for at least two species as demonstrated withthe dispersant ICE-based SSDs. While A. bahia generallyappears to be more sensitivity than M. beryllina (as shown hereand elsewhere10), SSDs from ICE models with M. beryllina asthe surrogate species did not result in significantly larger HCvalues. Consequently, the surrogate set of models that leads tosmaller HC values may be preferred, particularly whenprotection of especially sensitive species is a concern, orwhen there are concerns about species in specific microhabitat(e.g., sheltered salt marshes, mangroves or tidal flats) withincoastal ecosystems.One of the limitations of the current study is that, because of

the nature of the existing toxicity data, models were developedunder the assumption that petroleum hydrocarbons represent asingle compound with a predominant mode of toxicity(nonpolar narcosis). In reality, petroleum hydrocarbons are acomplex mixture of chemicals with more than one mode oftoxicity (narcosis, receptor-mediated). Because compounds inthese mixtures have different chemical properties and affinitiesfor lipids, it is widely recognized that their compositiondetermine their overall toxicity.43−45 This limitation, however,could be overcome by using quantitative structure activityrelationships (QSARs),2 which are relationships based onchemical structure. QSAR toxicity data can provide surrogatevalues for the development of hydrocarbon-specific ICEmodels,2 or could be used to develop QSAR surrogate-predicted species ICE models.46 Future refinements of ICEmodels are dependent upon greater availability of detailedanalytical chemistry results (e.g., individual PAH analytes),which are currently lacking in most petroleum and dispersanttoxicity data sources.13

In this study petroleum and dispersant ICE models weredeveloped and used to generate ICE-based SSDs. While thedevelopment of ICE models was limited by the availability oftoxicity data meeting the rigorous standardization criteria, theinformation presented here could facilitate assessments of thepotential toxicological consequences oil and dispersants toaquatic communities, aid in the estimation of concentrationassociated with low or no effects, and allow for comparisons ofthe relative sensitivity across test species. ICE-based SSDscould also be used in conjunction with environmental

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724570

concentrations from fate models or environmental monitoringto help characterize, via joint probability distribution curves, thefraction of potentially affected species or the magnitude ofadverse effects to aquatic communities.8 The inclusion ofadditional paired-empirical data for a wider number of species,particularly for sensitive life stages, may allow for furtherapplication of ICE models in damage assessments so as to allowcomparisons across communities (e.g., epibenthic vs benthiceffects assessments) and habitats (e.g., tidal flats vs marshes).While model verification showed promising results, additionaltoxicity data could help improve existing ICE models andfacilitate the development of additional ones. Of special interestis the inclusion of paired standard test-sensitive or rare speciestoxicity data, which is essential to refine petroleum anddispersant benchmark concentrations protective of the mostsensitive or untested species.

■ ASSOCIATED CONTENT

*S Supporting InformationAdditional information related to ICE models is available inSupporting Information 1. This information includes acomplete list of references of the original sources, One tablecontaining scientific and common names, two tables containingstatistically significant ICE model parameters, 1 Tablecomparing empirical and ICE based HC5s, 1 Figure of themodel building scheme, two figures of the fitted models, onefigure on taxonomic relatedness, and one figure withrepresentative SSDs. Supporting Information 2 contains coredata used in the development of ICE models. This material isavailable free of charge via the Internet at http://pubs.acs.org.

■ AUTHOR INFORMATION

Corresponding Author*Phone: +1 803 254 0278; fax: + 1 803 254 6445; e-mail:[email protected].

NotesThe authors declare no competing financial interest.

■ ACKNOWLEDGMENTS

Special thanks to C. Jackson and S. Raimondo (U.S. EPA) forcomments to an earlier version of this manuscript. Thisresearch was made possible by a grant from NOAA and theUniversity of New Hampshire’s Coastal Response ResearchCenter (Contract No. 13-034) to Research Planning, Inc. Noneof these results have been reviewed by CRRC and noendorsement should be inferred. The views expressed in thisarticle are those of the authors and do not necessarily reflect theviews or policies of the U.S. EPA. This publication does notconstitute an endorsement of any commercial product.

■ ABBREVIATIONS

adj-R2 adjusted coefficient of determinationHC1 and HC5 1st and fifth percentile hazard concentrationsICE interspecies correlation estimationLC50 and EC50 median lethal and effects concentrations,

respectivelyMSE mean square errorSSD species sensitivity distributionsTHC total hydrocarbon content

■ REFERENCES(1) Asfaw, A., Ellersieck, M. R., Mayer, F. L. Interspecies CorrelationEstimations (ICE) for Acute Toxicity to Aquatic Organisms and Wildlife.II. User Manual and Software, EPA/600/R-03/106; United StatesEnvironmental Protection Agency: Washington, DC, 2003.(2) Dyer, S. D.; Versteeg, D. J.; Belanger, S. E.; Chaney, J. G.; Mayer,F. L. Interspecies correlation estimates predict protective environ-mental concentrations. Environ. Sci. Technol. 2006, 40 (9), 3102−3111.(3) Golsteijn, L.; Hendriks, H. W.; van Zelm, R.; Ragas, A. M.;Huijbregts, M. A. Do interspecies correlation estimations increase thereliability of toxicity estimates for wildlife? Ecotoxicol. Environ. Saf.2012, 80, 238−243.(4) Raimondo, S.; Jackson, C. R.; Barron, M. G. Influence oftaxonomic relatedness and chemical mode of action in acuteinterspecies estimation models for aquatic species. Environ. Sci.Technol. 2010, 44 (19), 7711−7716.(5) Raimondo, S.; Mineau, P.; Barron, M. Estimation of chemicaltoxicity to wildlife species using interspecies correlation models.Environ. Sci. Technol. 2007, 41 (16), 5888−5894.(6) Dyer, S. D.; Versteeg, D. J.; Belanger, S. E.; Chaney, J. G.;Raimondo, S.; Barron, M. G. Comparison of species sensitivitydistributions derived from interspecies correlation models todistributions used to derive water quality criteria. Environ. Sci. Technol.2008, 42 (8), 3076−3083.(7) Awkerman, J. A.; Raimondo, S.; Barron, M. G. Development ofspecies sensitivity distributions for wildlife using interspecies toxicitycorrelation models. Environ. Sci. Technol. 2008, 42 (9), 3447−3452.(8) Posthuma, L.; Suter II, G. W.; Traas, T. P. Species SensitivityDistributions in Ecotoxicology; Lewis: Boca Raton, FL, 2002.(9) De Zwart, D., Observed regularities in species sensitivitydistributions for aquatic species. In Species Sensitivity Distributions inEcotoxicology; Posthuma, L., Suter, G. W., Traas, T. P., Ed. LewisPublishers: Boca Raton, FL, 2002; pp 133−154.(10) Barron, M. G.; Hemmer, M. J.; Jackson, C. R. Development ofaquatic toxicity benchmarks for oil products using species sensitivitydistributions. Integr. Environ. Assess. Manage. 2013, 9 (4), 610−615.(11) Bejarano, A. C.; Levine, E.; Mearns, A. Effectiveness andpotential ecological effects of offshore surface dispersant use duringthe Deepwater Horizon oil spill: A retrospective analysis of monitoringdata. Environ. Monit. Assess. 2013, 185, 10281−10295.(12) de Hoop, L.; Schipper, A. M.; Leuven, R. S.; Huijbregts, M. A.;Olsen, G. H.; Smit, M. G.; Hendriks, A. J. Sensitivity of polar andtemperate marine organisms to oil components. Environ. Sci. Technol.2011, 45 (20), 9017−9023.(13) Bejarano A. C.; Clark J. R.; Coelho, J. M., Issues and challengeswith oil toxicity data and implications for their use in decision making:A quantitative review. In Environ. Toxicol. Chem. DOI: 10.1002/etc.2501, 2014.(14) Aurand, D.; Coelho, G. Cooperative Aquatic Toxicity Testing ofDispersed Oil and the “Chemical Response to Oil Spills: Ecological EffectsResearch Forum (CROSERF)”. Technical Report.; EM&A Inc.: Lusby,MD, 2005; p 79.(15) Clark, J. R.; Bragin, G. E.; Febbo, R.; Letinski, D. J. In Toxicity ofPhysically and Chemically Dispersed Oils under Continuous andEnvironmentally Realistic Exposure Conditions: Applicability to DispersantUse Decisions in Spill Response Planning, Proceedings of the 2001International Oil spill Conference; American Petroleum Institute:Tampa, FL, March 26−29, 2001; pp 1249−1255.(16) Singer, M.; Smalheer, D.; Tjeerdema, R.; Martin, M. Toxicity ofan oil dispersant to the early life stages of four California marinespecies. Environ. Toxicol. Chem. 1990, 9 (11), 1387−1395.(17) Singer, M. M.; George, S.; Jacobson, S.; Lee, I.; Tjeerdema, R.S.; Sowby, M. L. Comparative toxicity of Corexit ® 7664 to the earlylife stages of four marine species. Arch. Environ. Contam. Toxicol. 1994,27 (1), 130−136.(18) Bejarano, A. C.; Chu, V.; Farr, J.; Dahlin, J., Development andapplication of DTox: A quantitative database of the toxicologicaleffects of dispersants and chemically dispersed oil. In Proceedings of the

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724571

2014 International Oil Spill Conference, American Petroleum Institute:Savanah, GA, 2014, in press.(19) USEPA, ECOTOX. Release 4.0. Ecotoxicology Database; USEnvironmental Protection Agency: Duluth, MN, 2012. www.epa.gov/ecotox.(20) Neter, J.; Wasserman, W.; Kutner, M. H., Applied LinearStatistical Models: Regression, Analysis of Variance, And ExperimentalDesigns; McGraw-Hill Higher Education: Boston, MA, 1990.(21) Piegorsch, W. W.; Bailer, A. J., Statistics for Environmental Biologyand Toxicology; Chapman & Hall: London, UK., 1997.(22) R: A Language and Environment for Statistical Computing,Version2.13.2011-09-30; R Development Core Team: Vienna, Austria, 2011.(23) Canty, A.; Ripley, B. boot: Bootstrap R (S-Plus) Functions, Rpackage version 1.3-9, 2013.(24) Davison, A. C.; Hinkley, D. V., Bootstrap Methods and TheirApplications; Cambridge University Press: Cambridge, 1997.(25) Bejarano, A. C.; Farr, J. K. Development of short acute exposurehazard estimates: A tool for assessing the effects of chemical spills inaquatic environments. Environ. Toxicol. Chem. 2013, 32 (8), 1918−1927.(26) Looney, S. W.; Gulledge, T. R., Jr Use of the correlationcoefficient with normal probability plots. Am. Stat. 1985, 39 (1), 75−79.(27) Bragin, G.; Clark, J.; Pace, C. Comparison of Physically andChemically Dispersed Crude Oil Toxicity to Both Regional and NationalTest Species under Continuous and Spiked Exposure Scenarios, MSRCTechnical Report Series 94-015; Marine Spill Response Corporation:Washington, DC, 1994; pp 1−45.(28) Bragin, G. E.; Clark, J. R. Chemically Dispersed Crude Oils:Toxicity to Regional and National Test Species under Constant and SpikedExposures; Marine Spill Response Corporation: Washington DC, 1995;pp 1−44.(29) Hemmer, M. J.; Barron, M. G.; Greene, R. M. Comparativetoxicity of eight oil dispersants, Louisiana sweet crude oil (LSC), andchemically dispersed LSC to two aquatic test species. Environ. Toxicol.Chem. 2011, 30 (10), 2244−2252.(30) Pace, C. B.; Clark, J. R.; Bragin, G. E. In Comparing Crude OilToxicity under Standard and Environmentally Realistic Exposures,Proceedings of the 1995 International Oil Spill Conference; AmericanPetroleum Institute: Washington, DC, February 27 to March 2, 1995,1995; pp 1003−1004.(31) Perkins, R. A.; Rhoton, S.; Behr-Andres, C. Toxicity of dispersedand undispersed, fresh and weathered oil to larvae of a cold-waterspecies, Tanner crab (C. bairdi), and standard warm-water test species.Cold Reg. Sci. Technol. 2003, 36 (1), 129−140.(32) Perkins, R. A.; Rhoton, S.; Behr-Andres, C. Comparative marinetoxicity testing: A cold-water species and standard warm-water testspecies exposed to crude oil and dispersant. Cold Reg. Sci. Technol.2005, 42 (3), 226−236.(33) Rhoton, S. Acute Toxicity of the Oil Dispersant Corexit 9500, andFresh and Weathered Alaska North Slope Crude Oil to the AlaskanTanner Crab (C. bairdi), Two Standard Test Species, and V. fischeri(Microtox Assay). MSc Thesis, Institute of Northern Engineering,University of Alaska: Fairbanks, AK 2000.(34) Wetzel, D.; Van Fleet, E. S. In Cooperative Studies on the Toxicityof Dispersants and Dispersed Oil to Marine Organisms: A 3-Year FloridaStudy, Proceedings of the 2001 International Oil spill Conference;American Petroleum Institute: Tampa, FL, March 26−29, 2001; pp1237−1241.(35) Gardiner, W.; Word, J.; Word, J.; Perkins, R.; McFarlin, K.;Hester, B.; Word, L.; Ray, C. The acute toxicity of chemically andphysically dispersed crude oil to key arctic species under arcticconditions during the open water season. Environ. Toxicol. Chem. 2013,32 (10), 2284−2300.(36) Singer, M.; George, S.; Lee, I.; Jacobson, S.; Weetman, L.;Blondina, G.; Tjeerdema, R.; Aurand, D.; Sowby, M. Effects ofdispersant treatment on the acute aquatic toxicity of petroleumhydrocarbons. Arch. Environ. Contam. Toxicol. 1998, 34 (2), 177−187.

(37) Singer, M. M.; Jacobson, S.; Tjeerdema, R. S.; Sowby, M. InAcute Effects of Fresh versus Weathered Oil to Marine Organisms:California Findings, Proceedings of the 2001 International Oil SpillConference; American Petroleum Institute, Tampa, FL, USA, March26−29, 2001; pp 1263−1268.(38) Carriger, J. F.; Barron, M. G., Minimizing risks from spilled oilto ecosystem services using influence diagrams: The DeepwaterHorizon spill response. Environ. Sci. Technol. 45, (18), 7631-7639.(39) Coelho, G.; Clark, J.; Aurand, D. Toxicity testing of dispersed oilrequires adherence to standardized protocols to assess potential realworld effects. Environ. Pollut. 2013, 177, 185−188.(40) Singer, M.; Tjeerdema, R.; Aurand, D.; Clark, J.; Sergy, G.;Sowby, M. In CROSERF: Toward a Standardization of Oil Spill CleanupAgent Ecological Effects Research, Arctic and marine oil spill programtechnical seminar, 1995; Ministry of Supply and Services: Canada,1995; pp 1263−1270.(41) Rand, G.; Wells, P.; McCarty, L. Chapter 1. Introduction toAquatic Toxicology. In Fundamentals of Aquatic Toxicology Effects,Environmental Fate, And Risk Assessment; Rand, G., Ed.; Taylor andFrancis Publishers: North Palm Beach, FL, 1995; pp 3−67.(42) Barron, M. G.; Wharton, S. R. Survey of methodologies fordeveloping media screening values for ecological risk assessment.Integr. Environ. Assess. Manage. 2005, 1 (4), 320−332.(43) Di Toro, D. M.; McGrath, J. A.; Hansen, D. J. Technical basisfor narcotic chemicals and polycyclic aromatic hydrocarbon criteria. I.Water and tissue. Environ. Toxicol. Chem. 2000, 19 (8), 1951−1970.(44) Di Toro, D. M.; McGrath, J. A.; Stubblefield, W. A. Predictingthe toxicity of neat and weathered crude oil: Toxic potential and thetoxicity of saturated mixtures. Environ. Toxicol. Chem. 2007, 26 (1),24−36.(45) USEPA Procedures for the Derivation of EquilibriumPartitioning Sediment Benchmarks (ESBs) for the Protection ofBenthic Organisms: PAH Mixtures. EPA-600-R-02-013. http://www.epa.gov/nheerl/download_files/publications/PAHESB.pdf (accessed8 March 2013).(46) Chicu, S.; Putz, M. Koln-Timisoara Molecular activity combinedmodels toward interspecies toxicity assessment. Int. J. Mol. Sci. 2009,10 (10), 4474−4497.

Environmental Science & Technology Article

dx.doi.org/10.1021/es500649v | Environ. Sci. Technol. 2014, 48, 4564−45724572