Application of Interspecies Correlation Estimation (ICE) Models and QSAR in Estimating Species Sensitivity to Pesticides

Abstract

Ecological risk assessment is challenged by the need to assess hazard to the diverse communities of organisms inhabiting aquatic and terrestrial systems. Computational approaches, such as Quantitative Structure Activity Relationships (QSAR) and Interspecies Correlation Estimation (ICE) models, are useful tools that provide estimates of acute toxicity where data are lacking or limited for ecological risk assessments (ERA). This review describes the technical basis of ICE models for use in pesticide ERA that may be used in conjunction with QSAR model estimates or surrogate species toxicity data and demonstrates the potential for improving hazard assessment. Validation and uncertainty analysis of ICE model predictions are summarized and used as guidance for selecting ICE models and evaluating toxicity predictions. A user-friendly web-based ICE modeling platform (Web-ICE) is described and demonstrated through case studies. Case studies include the development of Species Sensitivity Distributions generated from QSAR and ICE estimates, comparative sensitivity for a pesticide and its degradate, and application of ICE-estimated toxicity values for listed species assessments.

1. Introduction

Pesticides are a structurally diverse group of chemicals that are intentionally applied in the environment and are designed to have mechanism-based targeted biocidal activities and selective toxicity. The assessment of the environmental hazards and risks of pesticides is complex because of the diversity of species potentially affected and protection goals that can range from individual organisms to populations, ecosystem structure and function, and species interactions and behaviors [1]. One of the continuing challenges in pesticide risk assessment is determining chemical hazards to the diverse communities of organisms inhabiting aquatic and terrestrial systems [2]. Assessing the sensitivity of threatened and endangered (listed) species is additionally challenging because of restrictions on toxicity testing, the limited availability of organisms, and the lack of standardized culture and test methods [3]. One approach adopted globally has been to test a limited number of standard organisms that are assumed to represent the diversity of potentially affected species. The standard test species are then used as surrogates for communities or species of concern; however, they may not adequately represent the sensitivity of threatened, endangered, or other special status species [4].

Computational approaches are used internationally in pesticide hazard assessment to address the limitations of available data. One widely used approach is the development and application of quantitative structure activity relationship (QSAR) models [5]. QSARs have been widely adopted because they can estimate toxicity based only on physical-chemical properties and structural attributes. In aquatic toxicology, QSAR models are available for a broad range of chemicals but predict toxicity to only a few standard test organisms and do not address the broader range of taxa within aquatic communities [6]. A second approach that augments species sensitivity databases and predicts pesticide toxicity to specific taxa of organisms is the use of Interspecies Correlation Estimation (ICE) models developed from existing acute toxicity data. ICE models are log linear regressions of the toxicity values of compounds tested in a pair of taxa [7]. Extensive toxicity testing in the 1980s and 1990s included a diversity of species for thousands of compounds across a wide range of chemical classes and modes of action. Compilation of these data provide the basis for the development of ICE models across a broad taxonomic and chemical space, greatly increasing the diversity of species for which toxicity can be estimated.

While current-day in vivo testing is typically limited to a relatively few standard species or no test species at all, QSAR and ICE models can extrapolate sensitivity of untested species from standard surrogate species or from QSAR predicted toxicity. The developed ICE models can then be used to evaluate safety levels of a pesticide based on taxonomically diverse set of species [7,8]. The US EPA provides a suite of ICE models for use in pesticide risk assessment in the internet platform Web-ICE (www3.epa.gov/webice/) [9] that includes a diversity of surrogate and predicted species. This paper provides a review of the technical basis of models used in Web-ICE (Section 2), an overview of the Web-ICE platform (Section 3), previously applied ICE-based toxicity estimation for pesticides and linkage to QSAR (Section 4), and conclusions and future recommendations for ICE models in pesticide risk assessment (Section 5).

2. Technical basis of models used in Web-ICE

2.1. Databases

Three databases were used to develop ICE models for three different taxonomic groups: Aquatic animals (fish and crustaceans), algae, and wildlife (birds and mammals). Data were required to be consistent with standard acute toxicity test protocols outlined by the American Society for Testing and Materials [e.g., 10,11, and earlier versions] and the US Environmental Protection Agency (US EPA) Office of Prevention, Pesticides, and Toxic Substances [e.g., 12,13, and earlier versions). Each database was standardized to reduce extraneous variability in interspecies relationships based on characteristics of the tests for each group. Each record required a reported chemical name or structure with chemical active ingredient ≥ 90% and open-ended toxicity values (i.e. > 100 μg/L or <100 μg/L) were excluded. Synonymous chemical names occurring in the databases were identified by CAS number and grouped. Additional details pertaining to the development and standardization of each database are discussed thoroughly for aquatic animals in Raimondo et al. [7] and Willming et al. [14], for algae in Brill et al. [15], and for wildlife in Raimondo et al. [8], and are discussed briefly below. A summary of the number of records, chemicals, and species in each of the databases used in Web-ICE version 3.3 (release June 2016) is presented in Table 1. Approximately 450 pesticides were included throughout the databases.

Table 1.

A summary of the databases and models developed for aquatic animals, algae, and wildlife in Web-ICE version 3.3 [9].

Â			Attributes			Number of models
Database	Records	Species	Chemicals	Species	Genus	Family
Aquatic animals	8632	316	1499	1550	854	887
Algae	1647	69	457	58	44	0
Wildlife	4329	156	951	560	0	292

The aquatic animal toxicity database was compiled from the US EPA ECOTOXicology Knowledgebase System [ECOTOX; http://cfpub.epa.gov/ecotox; accessed date September 2014; 16], US EPA Office of Pesticide Programs ecotoxicity database, US EPA Office of Water Ambient Water Quality Criteria [17], US EPA OPPT Pre-Manufacture Notification (PMN), US EPA OPPT High Production Volume (HPV) Challenge Program, US EPA Office of Research and Development data sources [e.g. 18], the open literature, and targeted research [19, 20]. Data were 48h EC/LC50 for daphnids, midges, and mosquitoes and 96h EC/LC50 for fish and all other invertebrates. Only juvenile data for fishes, amphibians, insects, molluscs, and decapods were included, where all life stages were accepted for other taxonomic groups. Toxicity records for cadmium, chromium, copper, lead, nickel, silver, zinc, pentachlorophenol and ammonia were normalized according to Ambient Water Quality Criteria [17].

The algae database was compiled from ECOTOX (accessed May, 2010), Proctor and Gamble internal testing archives (accessed October 2012) and high quality US EPA data that were not included in ECOTOX (unpublished data available through November 2012). The data were tests of 72- and 96-hours durations and included effect concentrations relating to 50% of test organisms based on growth rate (E_rC50), biomass (E_bC50), or cell density (EC50). The validity and quality of each study was determined by evaluating its concordance to standard method guidelines such as those from the Organization for Economic Cooperation and Development [OECD; 21], the US EPA [12, 13], and the American Society for Testing and Materials [ASTM; 10, 11]. Algal taxon was required to be identified to at least genus level ensure consistent species or genus designations were ensured. For metals, toxicity records were grouped for each metallic cation, regardless of the type of salt or specific anion.

The wildlife dataset was comprised of single oral dose LD50 values collected from the open literature [22–26] and from datasets compiled by US EPA (e.g., Office of Pesticide Programs Ecotoxicity Database) and Environment Canada [27–28].

2.2. Model development and validation

Models were developed separately for each database and the following process was performed for each dataset prior to model development. Where multiple toxicity values were available for a single chemical and test organism, the range of values was determined. When that ranged exceeded a factor of 10, the data were reviewed for outliers. If obvious outliers were identified, they were excluded from model development. If no obvious outliers were noted, all records for that chemical and species were excluded from model development. For all other instances of multiple records per species and chemical, the geometric mean was calculated and used in model development. For genus and family level models, species geometric means were first calculated, followed by genus or family geometric means if there were multiple records for a single chemical.

ICE models were developed as least squares regression in which each species was paired with every other species by common chemical. Log-transformed toxicity values for each species and chemical were used in the models and a minimum of three common chemicals was required for each species pair. For each species pair, the ICE model is described as Log₁₀(Predicted Taxa Toxicity) = a + b*Log₁₀(Surrogate Species Toxicity), where a and b are the intercept and slope, respectively (Figure 1) [7,8]. For species level models, the predicted taxon was an identified species. For genus and family level models, the predicted taxon was comprised of all the species available within the higher taxonomic level. Models developed from the algae database were only at the species and genus level. The wildlife database only yielded models at the species and family levels. Only significant models (p-value < 0.05) were used for further analyses and included in Web-ICE. For algae that had multiple endpoints for a chemical (e.g., EC50, E_rC50, and E_bC50), data for the same effect metric was used in the model for both predicted and surrogate species with the same chemical whenever possible. Only when this was not possible were endpoint metrics mixed. The number of models developed for each database and taxonomic level of the predicted taxa are presented in Table 1.

Figure 1.

Example ICE model to predict sensitivity of Coho salmon (Oncorhynchus kisutch) to a chemical from the measured sensitivity of Rainbow trout (Oncorhynchus mykiss). Sensitivity is represented by Log10(LC50) in μg/L. MSE is the model mean square error and r2 is the coefficient of determination.

Models with a sample size of four or greater were validated using leave-one-out cross-validation. Here, each pair of acute values for surrogate and predicted species was systematically removed one at a time from the original model and a new model (submodel) was built with the remaining data. The removed value for the surrogate species was used to estimate the sensitivity of the predicted species from the submodel [7,8]. The N-fold difference of each predicted and measured value (non-transformed data) was used as a metric of prediction accuracy and was calculated as the maximum value of the estimated/measured or measured/estimated. The N-fold differences of all predicted toxicity values from aquatic animal models were used to relate model performance to taxonomic distance of the predicted and surrogate species and chemical mode of action [MOA; 7, 14]. Cross-validation data were also used to derive user guidance from model parameters as described in Section 2.3.

Taxonomic distance was assigned for each model such that surrogate and predicted species within the same genus = 1; family = 2; order = 3; class = 4; phylum = 5; kingdom = 6. Each cross-validated data point was assigned to one of the following prediction categories based on the N-fold difference of predicted and actual values: 5-fold (≤5-fold), 10-fold (>5-fold, ≤10-fold), 50-fold (>10-fold, ≤50-fold), and greater than 50-fold. The cross-validated data were compared among taxonomic distance and cross-validation categories using a Chi-square test for differences in probabilities [7]. Model prediction accuracy was significantly related to taxonomic distance, with 95% of removed datapoints predicted within 5-fold of the measured value for models developed for two species within the same genus. The percentage of cross-validated datapoints that were within the 5-fold prediction category decreased as the percentage of data points in the other prediction categories increased with increasing taxonomic distance [7, 14].

The MOA of chemicals used in models for fish and aquatic invertebrates were assigned based on the MOAtox classification scheme for acute toxicity [29]. MOAtox is a public domain dataset of 1213 chemicals that includes a diversity of metals, pesticides, and other organic compounds that encompass six broad and 31 specific MOAs. MOAtox classifications are based on a combination of high confidence MOA assignment approaches, including international consensus classifications, QSAR predictions, and weight of evidence professional judgment based on an assessment of structure and literature information [29]. Single MOA models were then developed for all possible species pairs and broad and specific MOA categories. Within data subsets of each MOA, every species was paired with every other species for model development and cross-validation as described above. Comparison of both the model mean square error (MSE) and prediction accuracy of MOA-specific models and models developed using all chemical data (all data models) showed that MOA-specific models were only more robust for species pairs in different phyla (e.g., fish and invertebrates; Figure 2).

Figure 2.

Comparison of ICE models using a) all data and b) data for chemicals with narcosis model of action for Daphnia magna and the Eastern oyster (Crassostrea virginica). Sensitivity is represented by Log10(LC/EC50) in μg/L. MSE is the model mean square error and r2 is the coefficient of determination.

2.3. Model user guidance

Model parameters (MSE, R², and slope) were evaluated against prediction accuracy using the cross-validation results in an iterative approach [14]. Here, values for MSE (ranging from 0.05 to 3.1), R² (ranging from 0.05 to 0.95), and slope (ranging from 0.1 to 6.0) were randomly selected. The submodels developed in the cross-validation process that contained MSE below the random value, R² above the random value and slope above the random value were identified. The percent of predicted values from these models that were within 5-fold of the actual value was determined. The process was reiterated until the combination of parameters that resulted in the highest percentage of data points predicted within 5-fold of the actual value was determined. This was performed separately for each taxonomic level for all models and for models with N ≥ 10 and N < 10. The highest MSE (0.95), lowest R² (0.6) and lowest slope (0.6) that corresponded to the highest percent accuracy were recommended for model selection guidance in Willming et al. [14]. Based on the results of this analysis, the taxonomic distance analysis described in Section 2.2, and professional judgement for confidence intervals, Rules of Thumb were assigned for model selection and prediction evaluation. They include:

1. Relatively low mean square error (MSE) (< 0.95)

2. High R2 value (> 0.6)

3. High slope (> 0.6)

4. Close taxonomic distance (< 4)

5. Narrow confidence intervals (one order of magnitude between lower and upper limit)

6. High cross validation success rate

3. Web-based Interspecies Correlation Estimation (Web-ICE)

The Web-ICE tool was developed by the US EPA to provide a suite of ICE models for toxicity extrapolation in hazard and ecological risk assessment [9]. The tool provides ICE models in a user-friendly internet platform and contains modules to: (1) calculate toxicity directly from a selected taxa model, (2) generate Species Sensitivity Distributions (SSDs) and compute hazard concentrations from available surrogate and predicted species, and (3) estimate the sensitivity of threatened and endangered species. Web-ICE was first available as an internet-based platform in 2007 [30] and has had subsequent updates in 2010 [31], 2013 [32], and 2016 [9]. Figure 3 shows the Web-ICE home page (v. 3.3), with modules for taxa-specific toxicity estimation (aquatic, wildlife, algae), generation of SSDs, and prediction of endangered species sensitivity. Links are also provided to user guidance and documentation of the technical basis of the tool.

Figure 3.

Home page of internet platform Web-ICE (version 3.3; accessed 2019 August 09).

For direct toxicity estimation, Web-ICE has separate modules for aquatic vertebrates and invertebrates at the species, genus, and family levels; aquatic algae at the species and genus levels; and wildlife at the species and family level. Each of these modules provides a complete set of model information and graphical representation of the model. The user enters the measured value of the surrogate and the module calculated the estimated toxicity and confidence intervals.

Web-ICE contains two SSD modules: one for aquatic and one for terrestrial wildlife species. The aquatic module can combine toxicity values for vertebrates, invertebrates, and algae. In each module, Web-ICE generates a log-logistic cumulative distribution functions of toxicity developed from simultaneously estimated toxicity values to all predicted species using up to 25 surrogates [33]. From the distribution, the concentration that represents the 1^st, 5^th, or 10^th percentile is calculated for use in hazard assessment. The module includes various functions for the user to filter and select predicted values to include or exclude in the distribution, and allows for export of the ICE-generated data into a spreadsheet format.

Endangered species modules are provided for terrestrial and aquatic animals to estimate toxicity values to listed species using all available models for up to 25 surrogate species. The aquatic endangered species module uses all available species, genus, or family level models while the terrestrial module uses only family level models. Users may predict to all available listed species within a broad taxonomic group (e.g., fishes) or to a particular species (e.g., Atlantic salmon, Salmo salar) using up to 25 surrogates. Toxicity estimations for listed species or their respective genera or family can also be exported into a spreadsheet format.

In addition to toxicity predictions, the Web-ICE interface allows users to download datasets that include all model information, chemicals and their MOAs (aquatic only) present within the models, and MOA-specific models for aquatic animals. A bibliography provides access to the extensive peer-reviewed research that provides the basis for the development and application of ICE models in hazard assessment.

4. ICE-based Toxicity Estimation for Pesticides

4.1. Species Sensitivity Distributions for Water Quality Criteria Development

Species sensitivity distributions (SSDs) are cumulative distribution functions of species toxicity values that can be used to compute a specified percentile hazard concentration that encompasses the sensitivity of taxa within the SSD [33]. The SSD approach has been utilized globally to develop ecological screening levels in ERAs and derive water quality criteria and standards for the protection of aquatic life [2, 34]. Broader application of the SSD approach to the hundreds of thousands of environmental contaminants and chemicals in commerce continues to be hindered by the limited diversity in species-specific toxicity values needed to compose SSDs [2].

ICE models and the Web-ICE platform provide a powerful tool to generate and augment SSDs using only limited toxicity data for aquatic and wildlife species, and ICE-based SSDs have been recommended for application in water quality criteria development as an alternative to generic safety factors for species extrapolation in international applications [35]. Awkerman et al. [36] developed ICE-based SSDs for pesticides, pharmaceuticals, and other chemicals and determined that ICE-estimated hazard levels were within 5-fold of 90% of measured values. Dyer et al. [37] compared aquatic toxicity hazard concentrations of 55 chemicals from the U.S. EPA Ambient Water Quality Criteria to those derived from ICE-generated values. Hazard levels from ICE-generated SSDs for pesticides, metals, and other compounds were shown to be within an average factor of 3.0 for multiple chemical classes that spanned over 7 orders of toxic magnitude [37]. A statistical assessment by Awkerman et al. [38] determined that augmentation of species diversity in SSDs with ICE predicted toxicity values provided robust estimates of hazard concentrations.

The Web-ICE tool (v3.3) contains a taxonomically rich suite of models to increase both sample size and taxonomic diversity in SSDs. Figure 4 illustrates the development of SSDs using measured (left, red) and ICE-generated (right, black) toxicity data for the organochlorine pesticide dieldrin as an example. Here, the fifth percentile hazard concentration (HC5) of dieldrin derived from only measured toxicity values was 0.684 mg/L, compared to an HC5 of 1.3 mg/L from an ICE-based SSD generated with a single measured toxicity value (Fig. 4). Since ICE models may be used to either generate entire SSDs from only a few surrogate species [e.g., 37] or augment an existing dataset to increase taxonomic diversity [38], variation of HC5s attributed to ICE models will depend on the extent to which the SSD is based on ICE data.

Figure 4.

Species Sensitivity Distributions for dieldrin using measured data (left, open circles) and ICE-generated data (right, black).

4.2. Direct comparison of pesticide and degradate toxicity using Web-ICE

Assessments of the ecological risks of pesticides can require determining the impacts of a compound throughout its life cycle. An assessment of the organochlorine insecticide endosulfan included determining whether compounds formed from environmental degradation of the active ingredient would be of equal or greater toxicity than the parent compound [39]. The available information for endosulfan and its primary degradate, endosulfan sulfate, showed measured toxicity of the two compounds within an order of magnitude for five aquatic species (bluegill, carp, sheepshead minnow, Daphnia magna, mysid; Table 2). For terrestrial species, the only measured toxicity data for endosulfan sulfate were for Northern bobwhite, which was not tested for endosulfan. Endosulfan data were available for the mallard duck and the rat (Table 3).

Table 2.

Comparison of the acute toxicity estimates for endosulfan and endosulfan sulphate using Web-ICE v 3.1 [31] and measured data for aquatic species.

Â	Â			Endosulfan sulphate
Species	Endosulfan (Âμg/L)	Toxicity (Âμg/L)	Web-ICE surrogate	Model level
Bluegill (Lepomis macrochirus)	1.7	3.8	NA	NA
Carp (Cyprinus carpio)	0.1	2.2	NA	NA
Daphnia magna	166	300	NA	NA
Sheepshead minnow (Cyprinodon variegatus)	1.3c	3.1	NA	NA
Mysid (Americamysis bahia)	0.83c	7.9	NA	NA
Rainbow trout (Oncorhynchus mykiss)	0.83	4.23 (3.32–5.39)	Bluegill	Species
Channel catfish (Ictalurus punctatus)	1.5	1.95 (0.77–4.92)	Carp	Family
Fathead minnow (Pimephales promelas)	1.5	2.89 (1.35–6.21)	Carp	Species
Scud (Gammarus lacustris)	5.8	30.15 (8.27–109.93)	Mysid	Genus
Stonefly (Pteronarcys californica)	2.3	2.23 (0.49–10.08)	Bluegill	Family
Striped mullet (Mugil cephalus)	0.38	3.06 (0.66–14.25)	Bluegill	Family
Pink shrimp (Farfantepenaeus duorarum)	0.04	1.03 (0.28–3.82)	Bluegill	Family
Grass shrimp (Palaemonetes pugio)	1.31	156.2 (90.76–268.71)	Daphnia magna	Genus
Eastern oyster (Crassostrea virginica)	0.45	8.55 (3.71–19.71)	Bluegill	Species

Table 3.

Comparison of the acute toxicity estimates for endosulfan and endosulfan sulphate using Web-ICE and measured data for wildlife.

Species	Endosulfan (mg/kg)	Endosulfan sulphate (mg/kg)
Northern bobwhite (Colinus virginianus)	31.2 (23.7–41.2)a	44
Mallard duck (Anas platyrhynchos)	28	52.9 (38.3–73.1)b
Rat (Rattus norvegicus)	10–40 (females-males)	63.6 (33.0–122.6)b

The Web-ICE platform (v. 3.0) was used to estimate toxicity of endosulfan sulfate for species for which a measured endosulfan toxicity value existed, but the degradate toxicity was lacking. In the case of terrestrial wildlife, it was also used to estimate the toxicity of endosulfan for northern bobwhite. To predict toxicity of endosulfan sulfate to the untested species, all species, genus, and family models were extracted for each potential surrogate and predicted species. User guidance for selecting and evaluating ICE models were used when multiple models existed for a predicted species (multiple surrogates, multiple taxonomic levels). Confidence intervals were used to confirm that the models with the lowest MSE, highest R², highest cross-validation success rate, and closes taxonomic distances yielded the most robust predictions. Using Web-ICE, endosulfan sulfate toxicity was estimated for 11 additional species, allowing for the two compounds to be compared between 17 across diverse taxa, rather than 5 species that were limited to standard test species. The results of the analysis demonstrate the relatively equal toxicity of endosulfan and endosulfan sulfate, with the degradate being slightly less toxic than the parent compound (Table 2 and 3). Based on less or equal toxicity of endosulfan and its degradate, a separate assessment for endosulfan sulfate was not warranted.

4.3. Application of Web-ICE in assessing pesticide risks to threatened and endangered species

Assessing risks of pesticides to listed species remains a widespread challenge [4, 40]. In a comprehensive review of pesticide risk assessments for listed species, the US National Academy of Science (NAS) recommended the use of ICE models to estimate acute toxicity values for listed species in place of safety factors [3]. In response to the NAS recommendations, EPA performed biological evaluations on malathion, diazinon, and chlorpyrifos as pilot chemicals to evaluate approaches highlighted in the NAS recommendations [41–43]. For these three chemicals, Web-ICE was used to predict acute toxicity values for federally-listed listed aquatic species based on toxicity data for test species that are typically available from registrant submissions required under the Federal Insecticide, Fungicide, and Rodenticide Act (FIFRA): the rainbow trout (Oncorhynchus mykiss), bluegill sunfish (Lepomis macrochirus), and fathead minnow (Pimephales promelas). In this analysis, the potential for Web-ICE to predict toxicity values for listed fish species from standard FIFRA test species using all available species, genus, and family level models was determined. The endangered species module in Web-ICE allows for this evaluation using a single data entry page, which generates a report for all possible listed species based on the entered surrogates. Here, we highlight the process as it was performed for chlorpyrifos and listed species in the family Salmonidae.

Using one or more of the FIFRA test organisms as the predictor species (i.e., O. mykiss chlorpyrifos LC50 = 8.24 μg/L; L. macrochirus chlorpyrifos LC50 = 3.02 μg/L; P. promelas chlorpyrifos LC50 = 212.31 μg/L), Web-ICE was used to estimate sensitivity for five listed salmonids: Apache trout (Oncorhynchus gilae), Atlantic salmon (Salmo salar), Chinook salmon (Oncorhynchus tshawytscha), Coho salmon (Oncorhynchus kisutch), and Cutthroat trout (Oncorhynchus clarkii). At the genus level, all three surrogate species could predict to the genera Oncorhynchus, Salmo, and Salvelinus, which represent 11, 2, and 1 listed species, respectively, including those represented by the species-level models listed above. All three species could also predict to the family Salmonidae. The rainbow trout model was selected as the best predictor of toxicity at the species, genus, and family level for salmonid species because it has the tightest confidence intervals, smallest MSE, highest R², and is the most taxonomically related (Table 4). The range of values predicted using rainbow trout was 4.18 ug/L (Apache trout, species level model) – 13.87 ug/L (Chinook salmon, species level model). The family level prediction for Salmonidae was 11.51 ug/L. Given the close range of all estimated values, 4.18 ug/L could be selected as a conservative estimate of acute salmonid sensitivity to chlorpyrifos.

Table 4.

Web-ICE (v 3.2, accessed October 2014) model predictions for salmon exposed to chlorpyrifos [43]. Models in bold were selected as the most appropriate model based on the User Guidance (section 2.3).

Predicted taxa	Model level	Surrogate	Estimated toxicity (Âμg/L)	Lower 95% confidence intervals	Upper 95% confidence intervals	Degrees of freedom (N-2)	r2	Mean square error (MSE)	Cross-validation success (%)
Apache trout	species	Fathead minnow	56.76	19.13	168.35	3	0.92	0.1	80
Apache trout	species	Rainbow trout	4.18	2.77	6.3	3	0.99	0	100
Atlantic salmon	species	Bluegill	0.786	0.274	2.25	9	0.96	0.08	100
Atlantic salmon	species	Fathead minnow	58.5	9.84	347.73	4	0.88	0.33	66.66
Atlantic salmon	species	Rainbow trout	7.64	3.71	15.72	11	0.95	0.13	84.61
Chinook salmon	species	Rainbow trout	13.87	6.68	28.77	6	0.97	0.06	100
Coho salmon	species	Bluegill	2.65	0.912	7.73	11	0.91	0.15	92.3
Coho salmon	species	Fathead minnow	81.4	25.07	264.32	6	0.8	0.28	75
Coho salmon	species	Rainbow trout	9.98	7.75	12.86	15	0.98	0.02	100
Cutthroat trout	species	Bluegill	9.99	3.44	28.97	30	0.69	0.38	68.75
Cutthroat trout	species	Fathead minnow	86.8	39.79	189.37	20	0.71	0.42	72.72
Cutthroat trout	species	Rainbow trout	10.44	6.85	15.92	32	0.93	0.08	97.05
Oncorhynchus	genus	Bluegill	3.5	2.71	4.51	310	0.88	0.22	90.38
Oncorhynchus	genus	Fathead minnow	113.6	78.65	164.08	83	0.83	0.36	83.52
Oncorhynchus	genus	Rainbow trout	10.99	8.29	14.59	45	0.95	0.07	97.87
Salmo	genus	Bluegill	1.19	0.597	2.38	17	0.95	0.09	100
Salmo	genus	Fathead minnow	49.23	10.26	236.23	13	0.52	0.93	73.33
Salmo	genus	Rainbow trout	8.68	5.94	12.67	23	0.96	0.08	96
Salvelinus	genus	Bluegill	3.67	1.41	9.54	33	0.77	0.28	82.85
Salvelinus	genus	Fathead minnow	68.42	30.88	151.61	21	0.7	0.33	86.95
Salvelinus	genus	Rainbow trout	12.12	7.57	19.42	37	0.88	0.16	92.3
Salmonidae	family	Bluegill	3.48	2.71	4.47	312	0.88	0.21	91.08
Salmonidae	family	Fathead minnow	108.57	77.88	151.36	85	0.85	0.3	87.35
Salmonidae	family	Rainbow trout	11.51	8.91	14.86	55	0.95	0.07	98.24

The results of analyses such as the one demonstrated here provides a range of sensitivity estimates for salmonids that could to be used in an assessment, rather than relying solely on the sensitivity of the available surrogate species. Basing hazard concentrations on a range within which the sensitivity of a listed species could reasonably be expected to occur is a scientifically defensible approach that provides considerably less uncertainty than the application of safety factors on surrogate species [3].

4.4. Linking QSAR and ICE models

While QSAR has shown great utility in chemical toxicity estimation, QSAR modeling approaches have not been adapted to determine the sensitivity of a diversity of species typically addressed in ecological risk assessments. Barron et al. [44] assessed whether QSAR and ICE models could be coupled and used to generate SSDs and HC5s with reasonable accuracy. Ten chemicals were selected that encompassed several modes of actions, chemical classes, and pesticidal classifications. Median lethal concentrations (LC50s) for a fish (P. promelas) and an invertebrate (D. magna) were estimated using three QSAR tools that employed different computational approaches: ECOSAR (Ecological Structure Activity Relationships), ASTER (Assessment Tools for the Evaluation of Risk), and TEST (Toxicity Estimation Software Tool). The QSAR estimates for each chemical were then used as input into Web-ICE to generate chemical specific SSDs of the same taxa composition and the chemical-specific HC5 (Figure 5).

Figure 5.

Comparison fifth percentile hazard concentrations (HC5) computed from SSDs of pesticides and other chemicals using measured data and those estimated from ICE using ASTER, ECOSAR, and TEST values.

The accuracy of the QSAR-ICE estimated HC5s were determined by comparison to measured HC5s developed from an independent dataset of experimental acute toxicity values for a diversity of aquatic species [44]. Estimated HC5s showed generally similar agreement with measured HC5s determined when species composition of the chemical specific SSDs were identical. The results indicated that coupling QSARs and ICE models provided a completely in silico approach to hazard estimation that encompassed larger species diversity. However, LC50 variability and species composition were large sources of variability in the in silico estimated HC5s. Barron et al. [44] concluded that additional research on linking QSAR and ICE models was needed to reduce uncertainty in HC5s and to develop fully in silico computational approaches for predicting species sensitivity. However, their analysis demonstrates where this approach is robust, and demonstrates variability that could be used to inform the application of uncertainty factors when used in a risk assessment.

5. Conclusions and future recommendations

Early suggestions that the relationship of chemical sensitivity tested in two species could be harnessed to inform risk assessment date back to the late 1970s [45–48]; however, the comprehensive development, validation, uncertainty analysis, and exploration of application of these models had not been conducted prior to the work described herein. The US EPA has conducted extensive research to provide validated models and demonstrated applicability for these in silico approaches in ERAs (Table 5). In addition to the applications described here, Web-ICE is used by US state environmental agencies to screen toxicity profiles [e.g., 49], by industries for ERAs, and is included in the syllabi of in environmental studies courses in US universities. Beyond the US, ICE models are for use in estimate toxicity thresholds [35, 50]. Between 2016 and 2018, Web-ICE averaged approximately 440 visits per month, with the most frequent visitors affiliated with government, industry, and academic sectors across the US, Europe, and Asia (source: Google Analytics, accessed 30 January 2019).

Table 5.

Chronological summary of US EPA articles and reports documenting the development, technical basis, and application of Web-ICE databases, models, and SSDs

Citation	Description	Category
Raimondo et al. 2007 [30]	Web-ICE v. 2.0 User guidance and database documentation	User Guidance
Raimondo et al. 2007 [8]	Development and technical basis of wildlife models	Wildlife
Dyer et al. 2008 [37]	Comparison of ICE and measured SSDs	Aquatic
Awkerman et al. 2008 [36]	Development and validation of wildlife SSDs	Wildlife
Raimondo et al. 2008 [4]	Validation of SSDs for listed species	Listed species
Raimondo et al. 2009 [49]	Evaluation of standardization approaches	Aquatic
Awkerman et al. 2009 [50]	Rodent toxicity data extrapolation to wildlife using ICE SSDs	Wildlife
Raimondo et al. 2010 [7]	Development, validation, and uncertainty analysis of aquatic ICE models	Aquatic
Raimondo et al. 2010 [31]	Web-ICE v. 3.1 User guidance and database documentation	User Guidance
Barron et al. 2012 [44]	Application of ICE models in QSAR based toxicity estimation	Aquatic
Barron et al. 2015 [29]	Development of mode of action scheme for Web-ICE	Aquatic
Feng et al. 2013 [35]	Application to water quality criteria and ERA	Aquatic
Raimondo et al. 2013 [32]	Web-ICE v. 3.2 User guidance and database documentation	User Guidance
Awkerman et al. 2014 [38]	Augmenting SSDs with ICE models	Aquatic
Bejarano and Barron 2014 [51]	ICE development for petroleum products	Aquatic
Brill et al. 2016 [15]	Development and validation of ICE algal models	Aquatic
Raimondo et al. 2015 [9]	Web-ICE v. 3.3 User guidance and database documentation	User Guidance
Willming et al. 2016 [14]	Development and validation of additional models for listed species	Listed species
Bejarano et al. 2017 [52]	Framework for selecting ICE models	Aquatic

Protecting broad communities with limited species data continues to be a significant challenged and source of uncertainty in ERA for pesticides and other environmental contaminants. The US National Academy of Sciences concluded that the use of safety factors is an indefensible approach to account for lack of data and recommended the use of ICE models in assessments for listed species and to supplement SSDs [3]. ICE models are a validated alternative to safety factors and their utility in ERA was been demonstrated through direct toxicity estimation and SSD development and augmentation [37, 38, 41–43]. High confidence in ICE estimations occurs when adhering to the model user guidance that has been determined through rigorous statistical analyses [7, 14]. Model attributes (e.g., taxonomic distance of the predicted and surrogate species, model parameters, cross-validation success rate) should be used in selecting models with low uncertainty and should be evaluated holistically. The ICE databases and models will continue to be updated through EPA’s Web-ICE platform to increase chemical and species diversity and to incorporate advances in interspecies toxicity estimation. Linking ICE and QSAR models holds great promise for improving the species diversity of in silico toxicity estimation using structure activity models.