Skip to main content
NCBI home page
EPA Author Manuscripts logoLink to EPA Author Manuscripts
. Author manuscript; available in PMC: 2024 Sep 28.
Published in final edited form as: Toxicol Sci. 2023 Sep 28;195(2):145–154. doi: 10.1093/toxsci/kfad072

Comparison of in silico, in vitro, and in vivo toxicity benchmarks suggests a role for ToxCast data in ecological hazard assessment

Christopher M Schaupp 1,*, Erin M Maloney 2, Kali Mattingly 3, Jennifer H Olker 4, Daniel L Villeneuve 4
PMCID: PMC11217893  NIHMSID: NIHMS1973130  PMID: 37490521

Abstract

Large repositories of in vitro bioactivity data such as US EPA’s Toxicity Forecaster (ToxCast), provide a wealth of publicly accessible toxicity information for thousands of chemicals. These data can be used to calculate point-of-departure (POD) estimates via concentration-response modeling that may serve as lower bound, protective estimates of in vivo effects. However, the data are predominantly based on mammalian models and discussions to date about their utility have largely focused on potential integration into human hazard assessment, rather than application to ecological risk assessment. The goal of the present study was to compare PODs based on (1) quantitative structure-activity relationships (QSAR), (2) the 5th centile of the activity concentration at cutoff (ACC) and (3) lower-bound cytotoxic burst (LCB) from ToxCast, with the distribution of in vivo PODs compiled in the Ecotoxicology Knowledgebase (ECOTOX). While overall correlation between ToxCast ACC5 and ECOTOX PODs for 649 chemicals was weak, there were significant associations among PODs based on LCB and ECOTOX, LCB and QSARs, and ECOTOX and QSARs. Certain classes of compounds showed moderate correlation across datasets (e.g., antimicrobials/disinfectants), while others, such as organophosphate insecticides, did not. Unsurprisingly, more precise classifications of the data based on ECOTOX effect and endpoint type (e.g., apical vs. biochemical; acute vs. chronic) had a significant effect on overall relationships. Results of this research help to define appropriate roles for data from new approach methodologies (NAMs) in chemical prioritization and screening of ecological hazards.

Keywords: Alternative toxicity testing, new approach methodologies, ecological risk assessment, high-throughput screening, quantitative structure-activity relationships

Introduction

Archetypal chemical risk assessment is comprised of four components: hazard identification, dose-response assessment, exposure assessment, and risk characterization (Klaassen, 2013). The intention underpinning this strategy is to sequentially gather as much relevant information as possible in each step before moving onto the next, which, in concept, aids in making thoroughly informed decisions straightforward and transparent while reducing contention. In practice, however, rigid adherence to this paradigm is problematic, given the sheer number of chemicals to assess and the time and cost constraints in creating a full toxicological profile for each chemical. As a result, chemical hazard information is available for only a small portion of existing chemicals, leaving toxicity data gaps for many thousands of compounds (Judson et al., 2008). Furthermore, because human health has historically driven chemical safety research, toxicity data insufficiencies are even more pronounced with respect to ecological risk assessment.

A variety of resources and tools have been developed to aid in addressing ecotoxicity information gaps. Two widely applied examples are the Ecological Structure Activity Relationships (ECOSAR) predictive model (de Haas et al., 2011) and the Ecotoxicology Knowledgebase (ECOTOX; Olker et al. 2022; https://cfpub.epa.gov/ecotox/). ECOSAR provides toxicity predictions for thousands of chemicals and is based on quantitative structure-activity relationships (QSARs), which utilize physicochemical properties (e.g., neutral organics, aliphatic amines, anilines, surfactant [surface-active] organic chemicals) to make computational assessments of aquatic toxicity. Since the 1980s, the USEPA has routinely used QSARs to predict the toxicity to aquatic organisms (fish, daphnids, algae) for new industrial chemicals lacking test data (Sanderson et al., 2003). ECOTOX complements ECOSAR data by providing a publicly accessible, curated repository of in vivo ecological toxicity data assembled from peer-reviewed literature for over 12,000 chemicals and species (Olker et al., 2022). The regulatory landscape, and perhaps more importantly, toxicity testing, has changed in some important ways over the last four decades since ECOTOX’s inception, and ECOTOX has adapted to incorporate new types of data, such as gene expression and molecular endpoints. Furthermore, new systematic review tools are now being implemented to assess the literature in a more thorough, transparent, resource-maximizing manner. Throughout its history, ECOTOX has been regularly used by risk assessors to extract ecological toxicity data which provide a basis for ecological risk assessments (particularly for pesticides) and the development of benchmarks such as Water Quality Criteria (WQC; USEPA, 1985).

While ECOTOX is an important open source of in vivo toxicity data, it currently only covers a portion of the broader chemical space and does not include in vitro data. Similarly, because QSARs use mathematical equations developed from empirical data, their applicability domains are limited to compounds with certain physicochemical properties and/or modes of action (e.g., non-polar narcosis; uncoupling; interactions with specific enzymes or receptors; Escher and Hermens, 2002). So-called new approach methodologies (NAMs) provide a means to both complement available in vivo toxicity data and to expand the scope of QSARs that may be developed (Ashauer and Jager, 2018). Interest in development and application of in vitro, high throughput approaches for assessing chemical bioactivity is intended to increase the pace and scale at which toxicological evaluation can occur, allow for systematic evaluation, reduce costs, and minimize reliance on animal studies (Krewski et al., 2010). However, to date, evaluation of the utility of in vitro, high throughput, NAMs has largely focused on potential integration into human hazard assessment. By comparison, interest and translational research in applying NAM data to ecological risk assessment has developed comparatively recently (Villeneuve et al., 2019).

Potential integration of alternative testing strategies into ecological risk assessment is not predicated on the idea that new data streams can replace traditional toxicity testing per se, but rather that data from these types of assays could help guide chemical prioritization for further in-depth evaluation (Thomas et al., 2019). Though ToxCast assays are based on mammalian cells and enzymes, previous work has demonstrated the utility of ToxCast data in an ecological setting at the screening level; notable examples include pairing bioactivity and environmental exposure data through exposure-activity ratios (EARs) to prioritize chemical assessment in Great Lakes tributaries (Blackwell et al., 2017; Corsi et al., 2019) and harnessing ToxCast data to derive screening level benchmarks for potential use by the EPA’s Office of Water in the absence of formal aquatic life criteria (Schaupp et al., 2022). These studies demonstrate that in vitro data from ToxCast can be used to generate points-of-departure (PODs; concentrations at which elevated toxicity from specific adverse responses are observed) protective of biological effect. Understanding the context(s) in which NAMs perform favorably relative to in vivo data, as well as where they perform unfavorably, is a crucial component for evaluating broader potential application to ecological risk assessment.

Given toxicity information gaps in ecological risk assessment in the context of the potential role of NAMs, our objective was to examine how PODs derived from in vivo ECOTOX data compared to PODs derived from in vitro ToxCast and in silico QSAR data for a large set (>500) of substances, with the idea that positive relationship(s) may support the derivation of NAM-based interim or screening values for and prioritizing evaluation of data-poor compounds. Toward this end, we assessed overall correlations between datasets. All chemicals within the extended ToxCast/ECOTOX/QSAR dataset were also separated by chemical use and mode-of-action (MOA) classes and ECOTOX data type to evaluate whether correlations were strengthened by grouping based on chemical use, endpoints, or taxa. Our results suggest that some ToxCast- and QSAR-derived PODs could be used for relative prioritization and/or screening in ecological risk assessment, and importantly, inform which chemicals are likely to elicit specific pathways of toxicity that may not be adequately covered by commonly used QSAR models or traditional guideline toxicity testing.

Methods

Chemical Space

To define the appropriate chemical space for this work, we began by comparing the ToxCast Phase III list of chemicals (4,584 compounds) with the ECOTOX Knowledgebase (December 2021 update; 12,425 chemicals from both aquatic and terrestrial species). Overall, 2,494 chemicals had overlapping entries in ECOTOX and ToxCast, with 2,305 having at least one aquatic ECOTOX record (Figure 1). Next, exclusionary criteria for ECOTOX records were applied to standardize and ensure only high-quality data were included (see Supplemental Material); following the ECOTOX filtering process, 672 potential chemicals remained. A final filtering step was performed to exclude chemicals which had undefined bioactivity entries in ToxCast (meaning biological activity was not observed in the assays tested; 23 chemicals fit this description), resulting in a final chemical space containing 649 compounds with data deemed suitable for deriving PODs from ToxCast and ECOTOX. For the purposes of this evaluation, the terms “POD” and “benchmark” are used interchangeably and do not necessarily reflect traditional risk-based definitions—they are not uniformly derived (some are based on minimum effect, some fifth percentile, etc.) and are general descriptors of screening level effect values derived from each data source. A full list of chemicals evaluated in this study and their associated PODs are included in the Supplementary Material.

Figure 1.

Figure 1.

Open in a new tab

Development of the chemical dataset. Following data extraction, a series of filtering criteria were followed to ensure appropriate data quality for benchmark derivation.

Chemical Classifications

Recognizing that relationships between datasets were likely to differ based on chemical groupings, we classified compounds primarily based on use. Chemical classes and subclasses were assigned progressively based on common Chemical Abstract Service Identification (CASID), using data collated from two main databases—the Great Lakes Restoration Initiative (GLRI) chemical database (https://clowder.edap-cluster.com/spaces/6180851fe4b07e4aef62b089) and the EPA’s CompTox Chemicals Dashboard (CCD; comptox.epa.gov/dashboard; accessed February 2022). Use annotations were available for approximately 60% of the 649 chemicals. For instances in which multiple uses were listed in the GLRI database, annotation was assigned according to primary use; for CCD entries, annotations based on product use categories (PUCs) were used (Isaacs et al., 2020). Compounds with no use annotation(s) in either GLRI or the CCD were classified manually through searches of various open-access databases (e.g., PubChem [pubchem.ncbi.nlm.nih.gov], DrugBank [Wishart et al. 2017], Wikipedia [www.wikipedia.org], Pesticide Properties Database [Lewis et al. 2016]). Pesticides comprised the largest portion of chemicals (68%), followed by multi-use/industrial substances (16%), pharmaceuticals and personal care products (PPCPs; 11%), antimicrobials/disinfectants (2%), and flame retardants (1%), with <2% remaining unclassified (Table 1). Because the final chemical space was dictated by available ECOTOX entries, the high proportion of pesticides largely reflects the historical emphasis for ecotoxicity testing placed on these compounds as part of registration. Compounds within the “pesticide” class were binned into the following subclasses: acaricides, fungicides, herbicides, insecticides, and rodenticides. Insecticides were classified further as to primary mechanism of action, i.e., carbamates, nicotinic acetylcholine receptor (nAChR) inhibitors, organochlorines (OCs), organophosphates (OPs), pyrethroids, and miscellaneous (insecticides which did not fall into the other four subclasses; Table 1; Supplementary Material II).

Table 1.

Summary of the chemical use classification scheme for the 649 chemicals examined.

Class Subclass Sub-subclass Number of chemicals
Antimicrobials/disinfectants 15
Flame Retardants 6
Multi-use/Industrial 104
Pesticides 439
Acaricides 13
Fungicides 90
Herbicides 190
Insecticides 136
Carbamates 18
nAChR inhibitors 7
Organochlorines 13
Organophosphates 50
Pyrethroids 21
Miscellaneous 6
Rodenticides 4
PPCPs 72
Unclassified 12
Open in a new tab

To contrast comparisons by chemical use, we also used a second, more broad classification strategy based on chemical MOA groupings included in TEST. Predictions of MOA were obtained using TEST v5.1 (US Environmental Protection Agency, 2020), which determines MOA based on the MOATox classification scheme, containing the following eight categories: acetylcholinesterase (AChE) inhibition; non-AChE inhibition; anticoagulation; nicotinic acetylcholine receptor (nAChR) agonism; narcosis; neurotoxicity; reactivity; and uncoupling of oxidative phosphorylation (uncoupler). Experimental MOAs were prioritized over predicted MOAs. MOA classifications were available for all but 53 chemicals, which were designated as “NA” (Supplemental Material III). Chemical composition by classification scheme for different ECOTOX groupings is included in Figures S21 and S22.

POD derivation

Filtering, quality assurance/quality control (QA/QC), and POD generation was carried out in R (r-project.org), using code available on GitHub at https://github.com/emmaloney/ToxCast_Benchmark_Comparison/blob/main/Classifications.R. General methods used to derive PODs are presented below, with further detail available in the Supplemental Information.

ToxCast

To derive ToxCast-based PODs, bioactivity data were extracted from the invitroDB (v3.4) module within the CompTox Chemicals Dashboard (accessed March 2022); two benchmarks were derived from these data. The first, ACC5, was derived by calculating the 5th centile from the distribution of activity concentration at cutoff (ACC) values which fell below the cytotoxic burst in ToxCast, as previously described (Schaupp et al., 2022). The second, the lower-bound cytotoxic burst (LCB), was internally calculated from a Z-score derived from 33 cytotoxicity-related assays within ToxCast (Judson et al., 2016). The cytotoxic burst represents a chemical-specific concentration at which general cell stress and cytotoxicity are observed (e.g., disruption of membrane integrity). Hence, the LCB functionally represents a concentration at which a non-specific “burst” of bioactivity and cytotoxicity is observed. Chemicals with LCB values of 1,000 μM were excluded from evaluation, as this corresponds to the highest test concentration within the ToxCast screening protocol, and likely does not represent an accurate cytotoxicity estimate for a particular chemical.

ECOTOX

Full records for all overlapping ToxCast Phase III chemicals were downloaded from ECOTOX. Certain poorly defined effects data were excluded (e.g., NOECs/NOELs at the highest concentration tested; Supplemental Material). Reported mean effect concentrations were preferentially used for benchmark derivation. When unavailable, mean concentrations were estimated using reported minimum and maximum effect concentrations. Effect values were then converted to a standard concentration (mg/L; ppm). Quantiles (Q1 = 0.25, Q3 = 0.75) and interquartile ranges (IQR) were calculated based on effect concentration distribution for each chemical; effect concentrations < q1 – (1.5 * IQR) or > q3 + (1.5*IQR) were treated as outliers and censored in the data. Final ECOTOX PODs were derived by taking the minimum effect concentration (minEC) for each chemical within the filtered dataset; these PODs represent a diverse suite of benchmarks representing different endpoints (see “Parsing ECOTOX data” below).

QSAR

Two programs maintained by the EPA provided QSARs for this study (both of which use ECOTOX data for model construction): ECOSAR (Version 2.2) and the Toxicity Estimation Software Tool (TEST, v5.1, www.epa.gov/chemical-research/toxicity-estimation-software-tool-test; [which also includes models based on rodent data]). Specific QSARs were obtained from CCD (TEST) and EPISuite (ECOSAR). Note that domains of applicability (i.e. species to which modeling was applicable) were not addressed beyond considerations taken by the CCD and ECOSAR teams. When available, acute POD estimates for fish (LC50), daphnids (LC50), and algae (EC50) were extracted from these models. If both models provided an acute benchmark for a particular chemical, the minimum QSAR was identified and used for POD comparisons. Subsequently, QSAR PODs were matched based on CASID and species to compare to ECOTOX and ToxCast data on a general taxonomic basis (i.e., fish/frogs, invertebrates, plants). Note that of the 649 chemicals with overlapping ECOTOX and ToxCast data, 584 had available QSARs.

Parsing ECOTOX data

A variety of endpoints and effect types are captured within ECOTOX records. We were interested in investigating how comparisons between PODs differed by data type. Accordingly, we grouped ECOTOX data in three main ways: effect type (Tier 1 [apical] vs. Tier 2 [biochemical]), exposure duration (acute vs. chronic), and taxonomy (by species). Tier 1 effects comprise apical endpoints, such as mortality, population, growth, reproduction, development, injury, physiology, intoxication, and morphology, while Tier 2 effects include cellular and molecular level measurements, such as gene expression, enzymatic activity, and protein binding. To delineate test type, we used TEST_TYPE_DESC from ECOTOX to separate testing into acute and chronic based on the following descriptors: Acute study = “acute” and “subacute”; Chronic study = “chronic,” “full life cycle,” “subchronic,” “generational,” and “partial life cycle.” TEST_TYPE_DESC is only populated if study authors report a specific test type, and many do not. Thus, if no information was listed in TEST_TYPE_DESC, a cutoff of seven days in the OBSERVED_DURATION field was used to delineate an acute from chronic study. There is no scientific consensus on what constitutes acute and chronic testing scenarios which, of course, is a function of the life span of the test species. Finally, species were assigned to three taxonomic groupings—vertebrates (fishes and frogs), invertebrates (e.g., Daphnia magna), and aquatic plants (e.g., Lemna). A full list of the ECOTOX data types assessed and their group designations is included in Supplemental Text.

Data analyses

All statistical analyses and figure generation were performed using R, with code available on GitHub, as previously noted. Given the wide range of linear benchmark values (in mg/L) and to improve normality of data distributions, PODs were log10-, log10(x + 10−5), or log(x + 10−6) transformed prior to further evaluation. Histograms and quantile-quantile (qq) plots were used to examine benchmark distributions within each POD type, and outliers were excluded from further analysis. To evaluate the relationships between the diverse POD types in the parsed and unparsed data (Figure 2, Supplementary Figures S1–S15), correlative analyses were carried out using the log transformed data. Pearson correlation coefficients were calculated using the ‘cor’ R function and the ‘cor.test’ function was used to analyze the correlation coefficient and obtain the significance level of the correlation. An α of 0.05 was used to define statistically significant correlative relationships. Regression lines were fitted to the data and comparative analyses were visualized using the ‘ggscatter’ function in the ‘ggpubr’ R package (https://cran.r-project.org/web/packages/ggpubr/index.html). The following datasets were compared: ACC5, ECOTOX, LCB, and QSAR. ECOTOX data were also separated by data type (Tier 1, Tier 2), exposure duration (acute, chronic), and taxonomy (aquatic invertebrates, aquatic invertebrates, aquatic plants). Binning of these data (and the chemical classifications) provide the basis for comparisons displayed in Figures 2 and 3, as well as the Supplemental Material. Qualitatively, highly significant relationships (i.e. p < 0.001) with corresponding ρ ≥ 0.3 and n ≥ 10 were considered noteworthy.

Figure 2.

Figure 2.

Open in a new tab

Correlogram displaying associations between different POD datasets. Square color reflects the Pearson correlation coefficient, while an “X” designates non-significance (p > 0.05).

Figure 3.

Figure 3.

Open in a new tab

Correlation plots comparing ToxCast lower-bound cytotoxic burst and ECOTOX PODs (overall) based on taxonomic classifications. Panel A displays ECOTOX data only from fish, B includes only invertebrate ECOTOX data, and C illustrates plant data from ECOTOX. Data were log10-transformed for comparisons, with log10 of cytotoxic burst plotted on the y-axis and log10 of the minimum ECOTOX POD plotted on the x-axis. Statistical p-values and coefficients (ρ) are based on Pearson correlation analysis (α < 0.05).

Results

From the three data sources used to derive PODs, ToxCast, ECOTOX, and QSAR, five general comparisons were made: 1) ACC5-ECOTOX, 2) LCB-ECOTOX, 3) ACC5-QSAR, 4) LCB-QSAR, and 5) ECOTOX-QSAR. Relationships were first examined using all data without additional parsing based on chemical use class, species group, effect type, etc. Faceting by chemical class was then performed within each of the five comparisons. Additional parsing into more specific subsets was only pursued if Pearson’s correlation for the overall dataset was significant (p < 0.05). An exhaustive discussion of correlations within these subsets is not included here, but full results are captured in the Supplementary Material.

ACC5-ECOTOX

It has been previously postulated that by including assays relevant to many vertebrate toxicity pathways, ToxCast data may be reasonable for generating lower bound estimates of in vivo adverse effect levels for risk-based prioritization in relation to human health hazards (Paul-Friedman, 2020). To determine whether ToxCast may be applied in a similar fashion as lower bound estimates for adverse ecological effect levels, we examined correlations between ToxCast-derived ACC5 and minEC from ECOTOX. Across the 649 chemicals for which this relationship was considered, there was little correlation between the two PODs (ρ = 0.07, p = 0.08, n = 649; Figure 2; Supplementary Figure S1). Given that ACC5 is derived from ACCs below the cytotoxic burst exclusively, we could hypothesize only chemicals with very specific modes of action would show correlation with ACC5—however, faceting by chemical use did not notably improve the relationships. Statistically significant correlations were detected for organophosphate pesticides (ρ = 0.29, p = 0.044, n = 50) and PPCPs (ρ = 0.27, p = 0.021, n = 71), but correlation coefficients were still comparatively low and derived ACC5 values were not consistently lower than minEC (Figure 4; Supplementary Figure S1). Parsing by MOA, a significant positive association was observed only for AChE inhibition (ρ = 0.31, p = 0.0071, n = 74; Figure S2). Thus, even for subclasses with significant correlations, ACC5 values do not consistently represent protective benchmarks, and given the relatively weak statistical significance of these relationships, the utility of ACC5 to predict ECOTOX benchmarks should be regarded as unreliable (with respect to the overall chemical set examined here).

Figure 4.

Figure 4.

Open in a new tab

Relative ecological POD comparisons between datasets. The number included in the middle (orange) bar represents the number of chemicals for which PODs derived from each respective source were within an order of magnitude (log1010). The bars to either side of the orange bar for each comparison represent the chemicals from that dataset which were lower than corresponding PODs from the opposing dataset by more than one order of magnitude. For example, in the first comparison (ECOTOX vs. QSAR), of 584 chemicals, 128 ECOTOX and QSAR PODs were within an order of magnitude, and ECOTOX PODs were lower relative to corresponding QSARs by more than a factor of 10 for 378 substances, while QSAR estimates were protective relative to ECOTOX benchmarks by more than one order of magnitude for 78 substances. Note that for the cytotoxic burst-ECOTOX comparison, the overall sample size is substantially reduced due to many of the chemicals having a cytotoxic burst concentration of 1000 μM (which were censored for this analysis, but are shown in a grey bar).

Cytotoxic burst-ECOTOX

The “cytotoxic burst” in ToxCast is presumed to be indicative of a burst in activity that occurs when concentrations of the test chemical reach levels where non-specific interactions with membranes and proteins start to cause a broad suite of generalized cytotoxicity and cellular stress (Judson et al., 2016). Similarly, “baseline toxicity” (a.k.a. narcosis) refers to acute toxicity to aquatic organisms elicited by organics through a nonspecific mode of action that disrupts lipid bilayer homeostasis (Escher and Schwarzenbach, 2002). Most narcotic compounds are structurally nonpolar/hydrophobic, but some polar compounds have also been shown to induce narcosis-like effects (Escher et al., 2011). As such, we might expect LCB to correlate reasonably well with minEC from ECOTOX for chemicals that act through baseline toxicity.

In contrast to ACC5, LCB and minEC showed a significant correlation across chemicals (ρ = 0.20, p < 0.001, n = 344; Figures 2 & 3; Supplementary Figure S4A). When parsed by chemical use class, there was a particularly strong positive relationship observed for antimicrobials/disinfectants (ρ = 0.69, p = 0.038, n = 9; Supplementary Figure S4B). Parsing by MOA revealed a strong positive correlation for narcotic chemicals (ρ = 0.28, p = 0.00068, n = 147), as well as those eliciting non-AChE inhibition (ρ = 0.56, p = 0.024, n = 26; Figure S5). Given the positive correlation overall, we separated ECOTOX data to investigate whether certain endpoint and/or effect types drove these results. As shown in Figure 3, significant associations between LCB and minEC were observed across taxa/trophic levels. Correlations were strongest in the aquatic vertebrates (ρ = 0.25, p < 0.0001, n = 302), but were considerable across invertebrates and plants as well (Figure 3; Figure S8). Analyzing by endpoint type revealed statistically significant correlations between Tier 1 ECOTOX data correlated with LCB (ρ = 0.21, p < 0.001, n = 332; Figure 2; Supplementary Figure S6A), but not between Tier 2 data and LCB (ρ = 0.1, p = 0.33, n = 93; Supplementary Figure S6B). A significant relationship between LCB and Tier 2 ECOTOX data was found for herbicides (ρ = 0.86, p = 0.0004), albeit from a small sample size (n = 12; Supplementary Figure S6D). That apical endpoints (Tier 1) would correlate with a cell-based and biochemical in vitro assays better than molecular ECOTOX endpoints (Tier 2) may initially seem counterintuitive since in vitro tests are generally not designed to recreate gross organismal endpoints. Relative to Tier 2 data, which ostensibly is more mechanistic in nature, a substantial portion of Tier 1 toxicity data reported in ECOTOX is attributable to baseline toxicity (Pavan et al., 2005). It is unsurprising then that LCB, which is effectively a benchmark for non-specific toxicity, would correlate with ECOTOX data from endpoints which largely reflect non-specific toxicity. In addition, these findings may also be partially attributable to a much larger chemical space for Tier 1 ECOTOX endpoints relative to Tier 2 (332 vs. 93, respectively).

Next, we separated ECOTOX data by exposure duration to compare acute and chronic benchmarks to LCB and found that acute ECOTOX PODs were significantly correlated with LCB (ρ = 0.24, p < 0.001, n = 333; Supplementary Figure S7A), as were chronic PODs, though the association with chronic data was not as strong (ρ = 0.18, p = 0.0097, n = 209; Figure 2; Supplementary Figure S7C). Acute in vivo ecotoxicity may be more likely to be associated with nonpolar narcosis, whereas chronic adverse effects often result from sub-lethal modes of action which impinge on specific pathways, so this finding was anticipated. Together, our analyses suggest Tier 1 and acute ECOTOX data are significantly aligned with LCB, albeit with substantial scatter across the data.

ACC5-QSAR

Given that the QSAR models considered in the present study are largely designed to predict baseline toxicity of non-polar narcotics and a few other reactive modes of action, and ACC5 is derived from values which represent targets related to specific mechanisms of toxicity, we hypothesized that correlations with ACC5 may be weaker than correlations between LCB and QSAR estimates. As with ECOTOX minEC, ACC5 correlated poorly with QSAR PODs (ρ = 0.05, p = 0.2, n = 584; Figures 2 & S10). A potentially important factor contributing to the poor overall relationship between ACC5 and QSAR was the spread of QSAR data, which spanned 15 over orders of magnitude in terms of LC50 across chemicals. In contrast, the ACC5 data were spread over eight orders of magnitude. When separated by chemical groups, a significant positive relationship was observed for rodenticides, albeit based on a small sample size (ρ = 0.97, p = 0.026, n =4; Figure S10, panel B). No significant correlations were seen when separated by MOA (Figure S11). Despite being designed to target mammalian physiology (vitamin K metabolism), many rodenticides are known to have high off-target toxicity in fish (Regnery et al., 2019). These unintended ecological effects were borne out in the QSAR data, where toxicity estimates for aquatic species, and not rodents, provided the minimum QSAR for the four rodenticides in this dataset (bromadiolone, chlorophacinone, pindone, and sodium fluoroacetate).

Cytotoxic burst-QSAR

Based on the logic above, we would expect a stronger correlation between QSAR estimates and LCB, since both benchmarks estimate non-specific toxicity. Indeed, among the broad comparisons we examined for this study, the correlation between LCBs and QSARs was the strongest (ρ = 0.32, p < 0.001, n = 311; Figure 2), despite the range of LCBs (< 1000 μM) being relatively compressed (most data within one order of magnitude), which typically makes it more difficult to find significant correlations.

Parsed by use class, four chemical groups displayed significant correlations between LCB and QSAR: herbicides (ρ = 0.41, p = 0.0002), miscellaneous pesticides (ρ = 0.61, p = 0.002), multi-use/industrial chemicals (ρ = 0.77, p = 9.3 × 10−8), and antimicrobials/disinfectants (ρ = 0.85, p = 0.007) (Figure S13B). Separation by MOA revealed notable, strong associations for LCB and QSAR in narcotic (ρ = 0.52, p = 2.9 × 10−11) and uncoupling chemical (ρ = 0.77, p = 7.7 × 10−5) groupings (Figure S14).

Because QSARs were based on minimum values from across species, we were interested in evaluating LCB-QSAR by taxonomic group. LCB and QSAR consistently correlated across the taxonomic groups examined, with aquatic vertebrates displaying the strongest correlation (ρ = 0.59, p < 0.001, n = 155), followed by aquatic invertebrates (ρ = 0.31, p < 0.001, n = 147) and plants (ρ = 0.7, p = 0.04, n = 9; Figure S15).

Detailed investigation of QSAR and LCB PODs by chemical also yielded an interesting insight. Seven of the ten chemicals with the largest difference between LCB and QSAR benchmarks were sulfur-containing pesticides for which alga was the organism providing the lowest QSAR value (data not shown; see comments under ‘PREFERRED_NAME’ column in Supplemental Material II). For all seven of these chemicals, the QSAR estimate was at least nine orders of magnitude higher than the LCB (and corresponding ECOTOX PODs aligned more closely with LCB), suggesting that ECOSAR and TEST models are insufficient for estimating aquatic toxicities of sulfur-containing pesticides. Literature review did not afford potential explanation(s) for such a discrepancy; future investigation into QSAR modeling of the toxicity of sulfur compounds may be warranted.

ECOTOX-QSAR

Our final broad comparison was between minEC and QSAR. A substantial portion (>75%) of the reporting in ECOTOX reflects nonpolar or polar narcosis (Pavan et al., 2005). In many cases, the minimum EC is based on an apical endpoint, such as mortality, which may be driven by non-specific toxicity, even if a more specific sub-lethal mode of action is possible. Acknowledging that QSARs effectively predict narcosis, we therefore anticipated good correlation between ECOTOX and QSARs. This prediction was corroborated by our analysis, which showed that ECOTOX PODs and QSAR were significantly correlated across 584 substances (ρ = 0.35, p < 0.001; Figure 2). In addition, positive associations reached statistical significance for five classes of chemicals: antimicrobials/disinfectants (ρ = 0.73, p = 0.01, n = 11), miscellaneous pesticides (ρ = 0.73, p = 8.7 × 10−7, n = 34), multi-use/industrial chemicals (ρ = 0.52, p = 1.7 × 10−7, n = 88), PPCPs (ρ = 0.3, p = 0.02, n = 62), and pyrethroids (ρ = 0.67, p = 0.001, n = 20) (Figure S16B). In addition, significant associations between overall ECOTOX PODs and QSAR estimates were observed for AChE inhibitors (ρ = 0.28, p = 0.019), narcotics (ρ = 0.24, p = 1.9 × 10−5), neurotoxicants (ρ = 0.65, p = 4.2 × 10−6), and reactive chemicals (ρ = 0.43, p = 0.031) (Figure S17).

Separating ECOTOX data by endpoint tiers and study type revealed Tier 1, but not Tier 2, ECOTOX PODs are significantly correlated with QSAR (ρ = 0.35, p < 0.001, n = 567 vs. ρ = 0.14, p = 0.11, n = 125, respectively). ECOSAR models are derived from aquatic toxicity experiments which focus on apical endpoints for chemicals which in many cases exhibit non-specific toxicity, therefore, QSARs aligning with ECOTOX Tier 1 data is consistent with expectation.

One of the most intriguing findings from our analysis was that both acute and chronic ECOTOX PODs were correlated with corresponding QSAR estimates (Figure 2). This was a curious finding, given that conventional in silico models are generally poor at predicting chronic toxicity (Gleeson et al., 2012). Herbicides, for which chronic ECOTOX PODs and QSARs were very strongly associated (ρ = 0.55, p = 8.1 × 10−8), comprised approximately 33% of the chemical space for chronic minEC-QSAR comparisons (80 of 245 total chemicals), potentially driving the overall association with chronic benchmarks. This result warrants further investigation to more fully elucidate the relationship between QSAR predictions and chronic ECOTOX PODs.

Broad Observations

Though statistically significant correlations were often observed between PODs from different datasets, no dataset provided reliably protective benchmarks across the chemicals examined (Figures 2 & 3). Our findings also indicate that correlation is not a good indicator of “protectiveness.” As an example, a statistically significant correlation between ECOTOX and QSAR benchmarks was observed, but the ECOTOX POD was more conservative (i.e., lower) for >80% (478 of 584) of chemicals examined (Figure 4). Similarly, when comparing cytotoxic burst and QSAR (ρ = 0.32, p < 0.001), the QSAR estimate was protective in 437 instances, compared to only 139 instances for the cytotoxic burst (Figure 4). ToxCast-based PODs were protective relative to ECOTOX PODs for ~50% of the compounds investigated (Figure 4). Unsurprisingly, the use of a particular POD may be contingent on the availability of other PODs, so a holistic approach using data from multiple sources (when possible) would be prudent.

We also examined the chemical composition of each ECOTOX category to identify any major differences by taxa (vertebrate, invertebrate, plant), study length (acute, chronic), or effect type (tier 1 vs. tier 2); both chemical use and MOA classifications were considered (Figures S21 & S22). Except for Tier 2 ECOTOX data, the composition is relatively consistent across all ECOTOX categories, with herbicides, fungicides, and multi-use constituting the most frequently observed chemical classes, and narcosis and AChE inhibition comprising the most common MOAs (excluding unclassified or “NA”). It is perhaps unsurprising that the chemical composition associated with Tier 2 endpoints differs from that for Tier 1. Tier 2 is largely composed of pathway-based or mechanistically-oriented endpoints. Investigators often measure such endpoints when they already have some a priori understanding of a putative MOA, and can use that to inform endpoint selection. Consequently, Tier 2 endpoints are less likely to be measured/reported for chemicals with unknown or undefined modes of action, beyond more systemic, non-specific, mechanisms such as narcosis. Nonetheless, these data suggest it is unlikely that chemical composition in the ECOTOX data is driving differences in correlation between POD types.

Discussion

The substantial knowledge gap between chemicals with ecotoxicity data and those which lack data has necessitated investigation into alternative sources of toxicity information (Ceger et al., 2022). Accordingly, we centered our analysis around an important question: what potential utility can data from mammalian NAMs serve within ecological risk assessment? As a means of addressing this question, toxicity benchmarks based on in silico, in vitro, and in vivo data were derived for over 500 compounds. To our knowledge, this was the first study to systematically compare PODs calculated from QSAR, ToxCast, and ECOTOX data for a large set of substances. We found that ACC5 correlates poorly with ECOTOX and QSAR POD estimates. In contrast, associations between LCB and ECOTOX, LCB and QSARs, and ECOTOX and QSARs were statistically significant. Though the correlation coefficients for these relationships (ρ) were all less than 0.5, similar efforts comparing in vitro and in vivo data from a human health perspective have found comparable correlations (Paul Friedman et al., 2022). Specific classes of chemicals, such as antimicrobials/disinfectants and narcotics, generally associated well across data sources and types, though none were consistently correlated across all datasets.

Given the difficulty in recapitulating in vivo dosimetry, metabolism, and physiology using in vitro or in silico approaches, there are many reasons which may explain a lack of correlation between ToxCast ACC5 and in vivo studies or QSAR. First, ECOTOX PODs are based on an amalgamation of data from different species and study types, whereas ToxCast data comprise short-term testing protocols for mostly human-derived (some murine) cells and bioassays. An additional consideration is that data “coverage” across chemicals is not consistent within the ECOTOX knowledgebase. Some chemicals have thousands of records, while others have much less information. Such discrepancy results from a combination of published literature availability and because certain chemicals have been prioritized for searches and review for inclusion in ECOTOX. Due to this data limitation, along with variability in study design and endpoint type, published toxicity data in ECOTOX are inherently “noisy,” which may limit confidence in in vivo PODs across compounds. Furthermore, the most sensitive active hits in ToxCast, which drive ACC5 derivation, can often represent adaptive responses to toxicant exposure (e.g., cytochrome P450 [CYP] induction), rather than clear toxic effect(s).

While many biochemical signaling networks are conserved across taxa at the molecular level, proteins important for adaptive responses to toxicants such as CYPs show considerable variability (LaLone et al., 2018). Accordingly, responses to chemical exposure(s) between mammals and aquatic organisms can differ significantly at both cellular and physiological levels (Ankley et al., 2016; LaLone et al., 2016). Moreover, defining precise biochemical “tipping points” for many of these adaptive responses (e.g., oxidative stress) is difficult. Test species, assay type, genetic background, and environmental conditions can have marked influence on such estimates.

ACC5 estimates were protective relative to in vivo ECOTOX PODs for approximately half of the substances examined and trends within chemical classes were inconsistent. Nevertheless, ToxCast data can potentially prove useful in ecological hazard assessment in that they provide mechanistic insight which other data sources often cannot when using cytotoxic burst and ACCs in tandem. Active hits far below the corresponding cytotoxic burst for a chemical may indicate specific mode(s)-of-action for further consideration in deriving effects-based PODs. Thus, the current analysis suggests a potentially important role for the ToxCast ACC5, not in terms of providing lower bound toxicity estimates, but in determining whether a more specific mode of action may be relevant.

Recognizing and defining limitations of an experimental approach is essential and can help provide insight into unexpected or “negative” results. In the context of this case study, inherent differences between datasets likely limited potential associations from the outset. One such consideration is that in vitro to in vivo extrapolation (IVIVE) was not included in our analysis. When performing IVIVE, accurately defining chemical concentrations by compartment is crucial. Some aquatic in vivo studies used to derive PODs from ECOTOX include analytical confirmation of tissue and water concentrations, but this is not consistently reported in the literature, and thus is not available for all records in the knowledgebase. ToxCast does not provide any analogous information (i.e. concentrations of test chemical measured in the well and/or cells within a test plate). Without these data, it would be difficult to reliably extrapolate concentrations from a ToxCast assay to water concentrations in an in vivo aquatic animal experiment, and as a result, we relied on nominal concentrations for the analysis. As a first pass inquiry though, we felt it reasonable to present comparisons between PODs with little additional modification(s) to the data (e.g., uncertainty factors, modifying factors to account for species-to-species differences, IVIVE, nominal water concentrations, etc.).

Despite some limitations, reasonable correlations between LCB, QSAR, and ECOTOX PODs, particularly for Tier 1 ECOTOX data, were observed. Narcotic, baseline effects constitute a significant portion (ECOTOX, QSAR) or all (LCB) of the chemical toxicity data from these sources, which may help explain why these PODs, but not ACC5 or benchmarks based on ECOTOX Tier 2 data (which in many instances reflect specific mechanisms of action), show positive associations. These findings were further bolstered by analysis using MOA classifications, which showed significant associations within narcosis groupings. Thus, our results lend credence to considerable literature suggesting that QSAR models provide reasonable estimates of baseline toxicity, a meaningful finding considering QSARs are the easiest PODs to generate and, notably, are already widely used for screening new chemicals (e.g., within the New Chemicals Division of the EPA’s Office of Pollution Prevention and Toxics). Additional work parsing compounds by mechanism(s) of action using a framework similar to that proposed by Kienzler et al. (2019), will shed further light on the relationships undertaken in the present study.

In summary, our analysis represents an important first step in evaluating the potential application of in vitro mammalian bioactivity data to ecological risk assessment. This study suggests that employing ToxCast data to replace traditional toxicity tests and define final ecological criteria or benchmarks is not currently possible, although it does support the potential use of ToxCast PODs for relative prioritization and/or screening. Compared to in vivo ECOTOX PODs, ACC5 are not consistently protective across chemicals, so there is limited quantitative value in using ToxCast PODs as direct surrogates for in vivo effects (Figure 4). However, given that the cytotoxic burst represents a benchmark for non-specific toxicity, a scenario in which ToxCast bioactivity data contain multiple active hits and an ACC5 well below the LCB may indicate that a specific mode of action is relevant to the chemical of interest; judicious application of this type of qualitative information could prove useful to risk assessors in guiding further hazard evaluation. In a similar vein, our findings indicate that QSAR appears to be a relatively good predictor of non-specific toxicity in vitro (using cytotoxic burst as a proxy). (Regression model development and assessment would aid in evaluating the values of such prediction, but was outside the scope of the work presented here.) Therefore, significant bioactivity below the LCB in ToxCast would suggest that QSAR model estimates may be insufficient to support hazard assessment for a given chemical. Consequently, with caution, in vivo experimentation could be tailored to investigate, or at the very least consider, potential mechanisms of action gleaned from ToxCast data. Given that QSARs are employed regularly in regulatory toxicology, having a tool to determine if specific mode(s) of action may be relevant in the evaluation of a new chemical (which traditional QSAR modeling would likely miss) would be valuable. A logical extension is that such a tool could also potentially contribute to weight-of-evidence indicating the likelihood of a particular compound to exhibit narcotic versus non-narcotic toxicity. As the production and introduction of chemicals into industry and the environment continues to increase, the relevance of developing alternative, resource-efficient strategies to assess ecotoxicology will continue to grow. The findings provided here help establish an important base on which to build future investigation toward this goal.

Supplementary Material

Supplement1
SM III
SM II

Acknowledgements

The authors thank Drs. Gary Ankley and Steve Corsi for comments on an earlier version of this manuscript. We also thank Dr. Todd Martin for his assistance generating TEST data.

Funding

US EPA Office of Research and Development, Chemical Safety for Sustainability National Research Program. The views expressed in this paper are those of the authors alone and do not necessarily reflect the views of the USEPA.

Footnotes

Data Links

R code for data analyses and figure generation can be accessed at the following location: https://github.com/emmaloney/ToxCast_Benchmark_Comparison/blob/main/Classifications.R.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Ankley GT, LaLone CA, Gray LE, Villeneuve DL, Hornung MW. Evaluation of the scientific underpinnings for identifying estrogenic chemicals in nonmammalian taxa using mammalian test systems. Environ Toxicol Chem. 2016. Nov;35(11):2806–2816. doi: 10.1002/etc.3456. Epub 2016 Jun 28. [DOI] [PubMed] [Google Scholar]
  2. Ashauer R, Jager T. Physiological modes of action across species and toxicants: the key to predictive ecotoxicology. Environ Sci Process Impacts. 2018. Jan 24;20(1):48–57. doi: 10.1039/c7em00328e. [DOI] [PubMed] [Google Scholar]
  3. Blackwell BR, Ankley GT, Corsi SR, DeCicco LA, Houck KA, Judson RS, Li S, Martin MT, Murphy E, Schroeder AL, Smith ER, Swintek J, Villeneuve DL. An “EAR” on Environmental Surveillance and Monitoring: A Case Study on the Use of Exposure-Activity Ratios (EARs) to Prioritize Sites, Chemicals, and Bioactivities of Concern in Great Lakes Waters. Environ Sci Technol. 2017. Aug 1;51(15):8713–8724. doi: 10.1021/acs.est.7b01613. Epub 2017 Jul 18. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Ceger P, Garcia-Reyero Vinas N, Allen D, Arnold E, Bloom R, Brennan JC, Clarke C, Eisenreich K, Fay K, Hamm J, Henry PFP, Horak K, Hunter W, Judkins D, Klein P, Kleinstreuer N, Koehrn K, LaLone CA, Laurenson JP, Leet JK, Lowit A, Lynn SG, Norberg-King T, Perkins EJ, Petersen EJ, Rattner BA, Sprankle CS, Steeger T, Warren JE, Winfield S, Odenkirchen E. Current ecotoxicity testing needs among selected U.S. federal agencies. Regul Toxicol Pharmacol. 2022. Aug;133:105195. doi: 10.1016/j.yrtph.2022.105195. Epub 2022 Jun 2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Corsi SR, De Cicco LA, Villeneuve DL, Blackwell BR, Fay KA, Ankley GT, Baldwin AK. Prioritizing chemicals of ecological concern in Great Lakes tributaries using high-throughput screening data and adverse outcome pathways. Sci Total Environ. 2019. Oct 10;686:995–1009. doi: 10.1016/j.scitotenv.2019.05.457. Epub 2019 Jun 5. [DOI] [PubMed] [Google Scholar]
  6. de Haas EM, Eikelboom T, Bouwman T. Internal and external validation of the long-term QSARs for neutral organics to fish from ECOSAR™. SAR QSAR Environ Res. 2011. Jul-Sep;22(5–6):545–59. doi: 10.1080/1062936X.2011.569949. Epub 2011 Jul 7. [DOI] [PubMed] [Google Scholar]
  7. Escher BI, Schwarzenbach R. Mechanistic studies on baseline toxicity and uncoupling of organic compounds as a basis for modeling effective membrane concentrations in aquatic organisms. Aquat Sci. 2002;64, 20–35. 10.1007/s00027-002-8052-2 [DOI] [Google Scholar]
  8. Escher BI, Hermens JL. Modes of action in ecotoxicology: their role in body burdens, species sensitivity, QSARs, and mixture effects. Environ Sci Technol. 2002. Oct 15;36(20):4201–17. doi: 10.1021/es015848h. [DOI] [PubMed] [Google Scholar]
  9. Escher BI, Ashauer R, Dyer S, Hermens JL, Lee JH, Leslie HA, Mayer P, Meador JP, Warne MS. Crucial role of mechanisms and modes of toxic action for understanding tissue residue toxicity and internal effect concentrations of organic chemicals. Integr Environ Assess Manag. 2011. Jan;7(1):28–49. doi: 10.1002/ieam.100. [DOI] [PubMed] [Google Scholar]
  10. Gleeson MP, Modi S, Bender A, Robinson RL, Kirchmair J, Promkatkaew M, Hannongbua S, Glen RC. The challenges involved in modeling toxicity data in silico: a review. Curr Pharm Des. 2012;18(9):1266–91. doi: 10.2174/138161212799436359. [DOI] [PubMed] [Google Scholar]
  11. Isaacs KK, Dionisio K, Phillips K, Bevington C, Egeghy P, Price PS. Establishing a system of consumer product use categories to support rapid modeling of human exposure. J Expo Sci Environ Epidemiol. 2020. Jan;30(1):171–183. doi: 10.1038/s41370-019-0187-5. Epub 2019 Nov 11. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Judson R, Richard A, Dix DJ, Houck K, Martin M, Kavlock R, Dellarco V, Henry T, Holderman T, Sayre P, Tan S, Carpenter T, Smith E. The toxicity data landscape for environmental chemicals. Environ Health Perspect. 2009. May;117(5):685–95. doi: 10.1289/ehp.0800168. Epub 2008 Dec 22. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Judson R, Houck K, Martin M, Richard AM, Knudsen TB, Shah I, Little S, Wambaugh J, Setzer RW, Kothiya P, Phuong J, Filer D, Smith D, Reif D, Rotroff D, Kleinstreuer N, Sipes N, Xia M, Huang R, Crofton K, Thomas RS. Analysis of the Effects of Cell Stress and Cytotoxicity on In Vitro Assay Activity Across a Diverse Chemical and Assay Space. Toxicol Sci. 2016. Oct;153(2):409. doi: 10.1093/toxsci/kfw148. Epub 2016 Sep 7. Erratum for: Toxicol Sci. 2016 Aug;152(2):323–39. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Kienzler A, Connors KA, Bonnell M, Barron MG, Beasley A, Inglis CG, Norberg-King TJ, Martin T, Sanderson H, Vallotton N, Wilson P, Embry MR. Mode of Action Classifications in the EnviroTox Database: Development and Implementation of a Consensus MOA Classification. Environ Toxicol Chem. 2019. Oct;38(10):2294–2304. doi: 10.1002/etc.4531. Epub 2019 Sep 5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Klaassen CD, editor. Casarett and Doull’s Toxicology: the Basic Science of Poisons. New York: :McGraw-Hill Education, 2013. [Google Scholar]
  16. Krewski D, Acosta D Jr, Andersen M, Anderson H, Bailar JC 3rd, Boekelheide K, Brent R, Charnley G, Cheung VG, Green S Jr, Kelsey KT, Kerkvliet NI, Li AA, McCray L, Meyer O, Patterson RD, Pennie W, Scala RA, Solomon GM, Stephens M, Yager J, Zeise L. Toxicity testing in the 21st century: a vision and a strategy. J Toxicol Environ Health B Crit Rev. 2010. Feb;13(2–4):51–138. doi: 10.1080/10937404.2010.483176. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. LaLone CA, Villeneuve DL, Lyons D, Helgen HW, Robinson SL, Swintek JA, Saari TW, Ankley GT. Editor’s Highlight: Sequence Alignment to Predict Across Species Susceptibility (SeqAPASS): A Web-Based Tool for Addressing the Challenges of Cross-Species Extrapolation of Chemical Toxicity. Toxicol Sci. 2016. Oct;153(2):228–45. doi: 10.1093/toxsci/kfw119. Epub 2016 Jun 30. [DOI] [PubMed] [Google Scholar]
  18. LaLone CA, Villeneuve DL, Doering JA, Blackwell BR, Transue TR, Simmons CW, Swintek J, Degitz SJ, Williams AJ, Ankley GT. (2018). Evidence for cross species extrapolation of mammalian-based high throughput screening assay results. Environmental Science and Technology, 52(23), 13960–13971. 10.1021/acs.est.8b04587. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Lewis KA, Tzilivakis J, Warner D, Green A. An international database for pesticide risk assessments and management. Human and Ecological Risk Assessment: An International Journal, 2016;22(4), 1050–1064. DOI: 10.1080/10807039.2015.1133242. [DOI] [Google Scholar]
  20. Pavan M, Worth AP, Netzeva TI. Preliminary analysis of an aquatic toxicity dataset and assessment of QSAR models for narcosis. 2005, JRC Report No. 21749 EN. Accessible from: https://publications.jrc.ec.europa.eu/repository/handle/JRC31399. [Google Scholar]
  21. Olker JH, Elonen CM, Pilli A, Anderson A, Kinziger B, Erickson S, Skopinski M, Pomplun A, LaLone CA, Russom CL, Hoff D. The ECOTOXicology Knowledgebase: A Curated Database of Ecologically Relevant Toxicity Tests to Support Environmental Research and Risk Assessment. Environ Toxicol Chem. 2022. Jun;41(6):1520–1539. doi: 10.1002/etc.5324. Epub 2022 Apr 26. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Paul Friedman K, Gagne M, Loo LH, Karamertzanis P, Netzeva T, Sobanski T, Franzosa JA, Richard AM, Lougee RR, Gissi A, Lee JJ, Angrish M, Dorne JL, Foster S, Raffaele K, Bahadori T, Gwinn MR, Lambert J, Whelan M, Rasenberg M, Barton-Maclaren T, Thomas RS. Utility of In Vitro Bioactivity as a Lower Bound Estimate of In Vivo Adverse Effect Levels and in Risk-Based Prioritization. Toxicol Sci. 2020. Jan 1;173(1):202–225. doi: 10.1093/toxsci/kfz201. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. R Core Team (2022). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/. [Google Scholar]
  24. Regnery J, Friesen A, Geduhn A, Göckener B, Kotthoff M, Parrhysius P, Petersohn E, Reifferscheid G, Schmolz E, Schulz RS, Schwarzbauer J, Brinke M. Rating the risks of anticoagulant rodenticides in the aquatic environment: a review. Environ Chem Lett. 2019;17:215–240. 10.1007/s10311-018-0788-6. [DOI] [Google Scholar]
  25. Sanderson H, Johnson DJ, Wilson CJ, Brain RA, Solomon KR. Probabilistic hazard assessment of environmentally occurring pharmaceuticals toxicity to fish, daphnids and algae by ECOSAR screening. Toxicol Lett. 2003. Oct 15;144(3):383–95. doi: 10.1016/s0378-4274(03)00257-1. [DOI] [PubMed] [Google Scholar]
  26. Schaupp CM, LaLone CA, Blackwell BR, Ankley GT, Villeneuve DL. Leveraging ToxCast data and protein sequence conservation to complement aquatic life criteria derivation. Integr Environ Assess Manag. 2022. Apr 7. doi: 10.1002/ieam.4617. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Sobus JR, Strynar M, Swank A, Tornero-Valez R, Ulrich EM, Villeneuve DL, Wambaugh JF, Wetmore BA, Williams AJ. The Next Generation Blueprint of Computational Toxicology at the U.S. Environmental Protection Agency. Toxicol Sci. 2019. Jun 1;169(2):317–332. doi: 10.1093/toxsci/kfz058. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Thomas RS, Bahadori T, Buckley TJ, Cowden J, Deisenroth C, Dionisio KL, Frithsen JB, Grulke CM, Gwinn MR, Harrill JA, Higuchi M, Houck KA, Hughes MF, Hunter ES, Isaacs KK, Judson RS, Knudsen TB, Lambert JC, Linnenbrink M, Martin TM, Newton SR, Padilla S, Patlewicz G, Paul-Friedman K, Phillips KA, Richard AM, Sams R, Shafer TJ, Setzer RW, Shah I, Simmons JE, Simmons SO, Singh A, Sobus JR, Strynar M, Swank A, Tornero-Valez R, Ulrich EM, Villeneuve DL, Wambaugh JF, Wetmore BA, Williams AJ. The Next Generation Blueprint of Computational Toxicology at the U.S. Environmental Protection Agency. Toxicol Sci. 2019. Jun 1;169(2):317–332. doi: 10.1093/toxsci/kfz058. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. United States Environmental Protection Agency (USEPA). 1985. Guidelines for deriving ambient aquatic life advisory concentrations (EPA/822/R-86/100). Office of Water. [Google Scholar]
  30. USEPA. 2020. User’s Guide for TEST (version 5.1) (Toxicity Estimation Software Tool): A Program to Estimate Toxicity from Molecular Structure. [Google Scholar]
  31. Villeneuve DL, Coady K, Escher BI, Mihaich E, Murphy CA, Schlekat T, Garcia-Reyero N. High-throughput screening and environmental risk assessment: State of the science and emerging applications. Environ Toxicol Chem. 2019. Jan;38(1):12–26. doi: 10.1002/etc.4315. Epub 2018 Dec 20. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Williams AJ, Grulke CM, Edwards J, McEachran AD, Mansouri K, Baker NC, Patlewicz G, Shah I, Wambaugh JF, Judson RS, Richard AM. The CompTox Chemistry Dashboard: a community data resource for environmental chemistry. J Cheminform. 2017. Nov 28;9(1):61. doi: 10.1186/s13321-017-0247-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Wishart DS, Feunang YD, Guo AC, Lo EJ, Marcu A, Grant JR, Sajed T, Johnson D, Li C, Sayeeda Z, Assempour N, Iynkkaran I, Liu Y, Maciejewski A, Gale N, Wilson A, Chin L, Cummings R, Le D, Pon A, Knox C, Wilson M. DrugBank 5.0: a major update to the DrugBank database for 2018. Nucleic Acids Res. 2018. Jan 4;46(D1):D1074–D1082. doi: 10.1093/nar/gkx1037. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplement1
NIHMS1973130-supplement-Supplement1.docx (51MB, docx)
SM III
NIHMS1973130-supplement-SM_III.xlsx (453.8KB, xlsx)
SM II
NIHMS1973130-supplement-SM_II.xlsx (59.2KB, xlsx)

RESOURCES