Evaluation of Complex Mixture Toxicity in the Milwaukee Estuary (WI, USA) using Whole Mixture and Component-Based Evaluation Methods

Abstract

Anthropogenic activities introduce complex mixtures into aquatic environments, necessitating the evaluation of mixture toxicity during ecological risk assessments. There are many new approach methodologies (NAMs) that can be used to complement traditional approaches for conducting mixture assessments. This study aimed to demonstrate how traditional approaches and NAMs can be integrated and employed for mixture evaluation in a target watershed. Assessments were carried out over two years (2017 – 2018) across 8 – 11 study sites in the Milwaukee Estuary (WI, USA). Whole mixtures were evaluated on a site-specific basis by deploying caged fathead minnows (Pimephales promelas) alongside composite samplers for 96-h and characterizing chemical composition, in vitro bioactivity of collected water samples, and in vivo effects in whole organisms. Chemicals were grouped based on structure/mode of action, bioactivity, and pharmacological actions. Priority chemicals and mixtures were identified by assessing contributions to cumulative toxicity units (maximum cumulative ratio analyses) and co-variance with measured effects (random forest regression). Whole mixture assessments identified specific target sites for further evaluation in the Milwaukee Estuary, including four sites impacted by industrial chemical/fuel/polycyclic aromatic hydrocarbon mixtures, four sites impacted by pharmaceutical mixtures, and three low priority sites. Component-based and predictive analyses identified fourteen mixtures and sixteen chemicals which significantly contributed to cumulative effects, representing high and medium priority targets for further ecotoxicological evaluation, monitoring, or regulatory assessment. Overall, this study represents an important complement to single-chemical prioritizations, providing a more comprehensive evaluation of mixture effects of chemicals detected in a target watershed. Furthermore, it demonstrates diverse tools and techniques that can be employed and adapted for future mixture risk assessments in aquatic environments.

1. Introduction

Environmental monitoring studies have demonstrated that agriculture, industrial operations, and urbanization result in the unintentional introduction of chemical mixtures into aquatic environments. Recent surveys have reported complex mixture detection in aquatic environments across North America (Baldwin et al. 2016; Blackwell et al. 2019), Europe (Malaj et al. 2014; Beckers et al. 2018), and Asia (Nkoom et al. 2018; Peng et al. 2018). This is an issue of ecotoxicological concern as co-exposure to multiple compounds can cause cumulative effects, resulting in environmental risks not adequately captured in single-chemical assessments (Kortenkamp et al. 2019). Thus, there is a need to more routinely move beyond single chemical evaluations to determine how mixture exposures may impact aquatic biota (Drakvik et al. 2020).

A diverse range of techniques have been employed to evaluate or estimate the cumulative effect(s) of chemical mixtures. Broadly, mixture assessment techniques either tend to assess hazards on a holistic (whole mixture) or an individual component basis (ECETOC 2011; European Commission 2012). Holistic approaches (e.g., whole effluent toxicity testing, in situ effects evaluation, direct toxicity assessments, etc.) characterize the ecotoxicity of a whole mixture (including measured and unmeasured constituents), providing generalized assessments of cumulative biological effects (Boobis et al. 2011; ECETOC 2011). Alternatively, component-based approaches (e.g., toxicity quotient (TQ) models, ecotoxicological thresholds of concern (ecoTTC), maximum cumulative ratios (MCR), etc. (Price and Han 2011; European Commission 2012; Barron et al. 2021)) characterize the contributions and interactions of a defined set of individual mixture constituents, providing detailed assessments of which chemicals are likely to most significantly contribute to cumulative toxicity (European Commission 2012). Although these two general mixture approaches differ, they can yield complementary assessments of potential effects. Thus, integration of these approaches can yield valuable insights into the biological and chemical aspects of complex mixture toxicity.

Over the past few years, there have been significant advances in the fields of ecotoxicology and environmental risk assessment, yielding a range of novel assays, bioinformatics tools and databases, sometimes collectively referred to as ‘new approach methodologies’ (NAMs), that can be employed in complex mixture assessment (Kienzler et al. 2016). For example, integrated effects-based monitoring techniques (e.g., in situ caged organism deployments combined with in vitro screening assays and analytical chemical quantification) can identify important bioactivities and chemical constituents within the context of environmentally relevant, whole mixture exposures (Cavallin et al. 2020; Ankley et al. 2021). High-throughput in vitro screening tools (e.g., ToxCast^™ and Tox21 programs (Tice et al. 2013; United States Environmental Protection Agency 2018) and non-targeted biological approaches (e.g., metabolomics, transcriptomics, proteomics (Garcia-Reyero and Perkins 2011; Altenburger et al. 2012; Ekman et al. 2013) can aid in identifying chemicals/mixtures that may perturb biological pathways controlling higher-level apical responses (Ankley et al. 2010; Schroeder et al. 2016; Corsi et al. 2019). Finally, publicly accessible informatic resources (e.g., the web-based datahub CompTox Chemicals Dashboard (Williams et al. 2017) and other open-source tools (e.g., the Toxicity Estimation Software Tool (TEST) quantitative structure activity relationship (QSAR) model suite (US EPA 2016), collaborative Adverse Outcome Pathway Wiki (AOP-Wiki) (Society for Advancement of AOPs 2022) can provide data on putative mechanisms of action and/or ecotoxicological outcomes associated with single chemical exposure, improving our ability to evaluate or estimate mixture effects in target biota. Overall, these NAMs can supplement traditional mixture assessment approaches, allowing for an improvement of the biological depth and chemical coverage within complex mixture evaluations.

To demonstrate how traditional mixture assessment approaches and NAMs can be integrated and employed for complex mixture characterization, we carried out a study of mixture toxicity in the Milwaukee Estuary Area of Concern (Milwaukee, WI). The Milwaukee Estuary is located across a mixed land-use area, primarily defined by urban, agricultural, and industrial activity. This watershed receives pollution from both point and non-point sources, including historic industrial discharges, combined wastewater treatment plants, combined sewer overflows, and agricultural and urban run-off (US EPA 2020). As such, it represents a suitable target region for the evaluation of complex mixture toxicity. Multiple sites in the Milwaukee Estuary were evaluated over the course of two years (2017 – 2018) using a diverse suite of holistic and constituent-based techniques. Whole mixture toxicity was assessed in situ by deploying caged adult fathead minnows (Pimephales promelas) alongside composite autosamplers at a range of study sites and evaluating endocrine- and xenobiotic-related effects. Component-based assessments of complex mixture toxicity were carried out for each study site using analytical data and selected computational tools to estimate apical, non-apical, and pharmacological mixture effects. Finally, random forest regression was employed to bridge in situ and constituent-based mixture analyses, yielding individual chemicals and mixture groups that were important predictors of significant endocrine and xenobiotic response related effects measured at study sites. This study had two main objectives: 1) use whole-mixture toxicity evaluations (e.g., chemical detection characteristics and biological effect occurrence trends) to prioritize sites in the Milwaukee Estuary; and 2) employ constituent-based assessments and predictive analyses to prioritize individual contaminants and chemical mixtures detected at study sites. Through these objectives, this study aimed to identify site, chemical, and mixture targets for further evaluation on the basis of ecotoxicological potential and field-monitored effects. Furthermore, by using traditional methods alongside more novel strategies for mixture evaluation, this study endeavoured to demonstrate how integrating NAMs and classical techniques can provide a more comprehensive assessment of mixture toxicity in a target watershed.

2. Methods

2.1. Milwaukee Estuary Study Sites

In situ caged-fish studies were carried out in the Milwaukee Estuary (Milwaukee, WI) in 2017 (6/8/17 – 6/15/17) and 2018 (5/9/18 – 5/17/18) (Figure S1). All study sites were located near United States Geological Services (USGS) monitoring stations, allowing for the collection of site-specific hydrological and land-use data. Selection of study sites reflected the diverse sources of contaminants within the Milwaukee Estuary watershed. In 2017, caged fish were deployed at 8 sites across areas defined by industrial and urban land-use (Kinnikinic River (KKL), Menomonee River at 25^th Street (MET), Menomonee River near Church St. (MEC), and Underwood Creek (UCJ), a mix of industrial, urban, and agricultural land-use (Milwaukee River at the Mouth of Lake Michigan (MIM), and Milwaukee River at Walnut Street (MIP)), and agricultural/natural land-use (Milwaukee River at Estabrook (MIE), and Menomonee River at Freistadt Road (MEF)) In 2018, fish were deployed at 3 additional sites located downstream of wastewater treatment plants (WWTPs; the Jones Island Water Reclamation Facility (Milwaukee Outer Harbor – Jones Island Plume (JIP)), the Cedarburg Water Recycling Centre (Cedar Creek at Cedarburg (CCM)), and the Village of Newburg Sanitary Sewer Treatment Facility (Milwaukee near Newburg (MIN)). The control site was at the Great Lakes Toxicology and Ecology Division (GLTED; Duluth, MN), however the abbreviation MED (Midcontinent Ecology Division) is generally used to refer to the control site throughout this manuscript for consistency with site labeling at the time the study was conducted. Site details, including location and land-use characteristics are available in Table S1.

2.2. Caged Fish Deployment and Laboratory Studies

In situ and laboratory-based exposures were carried out using reproductively mature (6 – 7-month-old) fathead minnows obtained from the on-site culture facility at GLTED. Detailed methodology is described elsewhere (Kahl et al. 2014; Ankley et al. 2021). Briefly, fathead minnows were shipped to the study locale (overnight delivery service). Fish were acclimated to site-specific water temperatures and then transferred into cages (n = 2 cages per site; 6 male and 6 female fish per cage [total of 24]) accompanied by an automated composite water sampler (Kahl et al. 2014; composite sampler, manufactured in-house) and suspended close to mid-water column depths at each site (approximate depths listed in Table S1). Deployments were carried out for 96-h, and the automated samplers were timed to collect water every 10 minutes over the full duration to provide a time-integrated composite. Due to logistics associated with fish retrieval and sample collection and preservation, cage deployments were staggered so a limited number of sites were evaluated each day (typically 2 – 3 sites/day of 8 – 11 total, depending on the study year). Laboratory control studies were carried out at the GLTED laboratory. As with field deployments, fish were selected from cultures, stored in a cooler overnight and then transferred into tanks (n = 2 tanks per treatment; 6 male and 6 female fish per tank [total of 24]), and exposed to UV-treated, filtered Lake Superior water for 96-h (16:8 light-dark photoperiod). Exposures were carried out in 20-L tanks using 10-L of water with a continuous-flow test system (45 mL / min, 6.48 volume additions/day), and fish were fed thawed frozen adult brine shrimp daily (to satiation). All experimental procedures using fish were reviewed and approved by the Animal Care and Use Committee, in accordance with the Animal Welfare Act and Interagency Research Animal Committee guidelines.

2.3. Sample Collection and Fish Processing

Following the 96-h fish deployments, composite (auto-sampler) or grab water samples (~ 8 L / site) were at collected from each study site, placed in a cooler on ice, and transported to the University of Wisconsin Milwaukee (Milwaukee, WI) for further processing. Subsamples (1.2 – 1.4 L) were transferred into pre-cleaned amber glass bottles and shipped overnight to the USGS National Water Quality Laboratory (NWQL, Denver CO) for organic contaminant analysis. The remaining water was transported back to the GLTED on ice and stored at −20°C until further use. Fish were transported from the field to the lab in buckets of site water (~ 12 L; separate buckets for each cage), assessed for survival, and anesthetized/euthanized using buffered tricaine methanesulfonate (Finquel; Argent, Redmond, WA, USA). Wet weights were recorded, fish were dissected, and gonads, plasma, liver, and intestinal tissues were collected. Gonads were weighed to determine gonadosomatic indices, and all tissues were snap frozen (using liquid nitrogen or dry ice) and stored at −80°C until further use. To minimize potential variability, fish were sampled one location at a time in the same order they were retrieved from the field and RNAseZap (Ambion, Austin, TX) was used to clean equipment between sample processing. As fish processing occurred over the course of several days, collected tissues were stored at the University of Wisconsin Milwaukee at −80°C, and then transported on dry ice to GLTED where they were held at −80°C until further analysis.

2.4. Water Quality and Chemical Detection

2.4.1. Water Quality Characterization

Water quality was measured over the course of all in situ caged-fish deployments and laboratory-based exposures. Dissolved oxygen (DO; mg/L), conductivity (μS/cm), and pH were measured at the beginning (cage deployment) and end (fish collection) of each caged fish study, and daily during laboratory control studies. Temperature (°C) was measured every 30 min in both field- and laboratory-based studies using in situ HOBO^® temperature data loggers (Onset Computer Corporation, Bourne, MA). Hardness (mg/L as CaCO₃), alkalinity (mg/L as CaCO₃), and nutrient concentrations (total nitrogen (N), total phosphorous (P), nitrate (NO₃), phosphate (PO₄), and ammonium (NH₄); μg/L) were evaluated in composite or grab water samples. Basic water chemistry data are available in the Supplemental Information (Table S2).

2.4.2. Chemical Analysis

Organic contaminant analysis was carried out at the USGS NWQL (Denver, CO, USA). Following protocols described in the NWQL laboratory schedule (LS) 4433, concentrations of wastewater indicators (e.g., wastewater contaminants, PAHs, and pesticides) were quantified in whole water samples via continuous liquid-liquid extraction with dichloromethane solvent followed by capillary column gas spectrometry/mass spectrometry (GC-MS/MS) (Zaugg et al. 2006). Following protocols described in the NQWL LS 2440, concentrations of pharmaceuticals in filtered water samples were quantified via liquid chromatography paired with tandem mass spectrometry (LC-MS/MS) (direct injection; positive ion mode). Further analytical methodology are available elsewhere (Zaugg et al. 2006; Furlong et al. 2014; Lee et al. 2015). For chemicals analyzed under both analytical schedules (atrazine, caffeine, cotinine), results from the method with the lowest reporting limit were used in analyses (NQWL LS 2440). In most cases, analytical measurements were carried out using composite water samples collected directly from study sites. However, grab samples were occasionally used when composite sample collection was unsuccessful (Table S3). Furthermore, in 2017 there were shipping issues associated with samples collected from four sites (MIM - field blank, MEC – technical duplicate, MEC, and KKL). Thus, for these sites analytical evaluation was carried out using reserve grab/composite samples stored at GLTED. Concentrations in technical duplicates were averaged by site and concentrations of organic contaminants in all study sites were corrected by subtracting concentrations of chemicals detected in field-blanks from mean/measured values. Concentrations reported as estimated values (E) were categorized as detects for ecotoxicological analyses. Concentrations categorized as detected but unquantifiable (M) were replaced with ¼ of the reporting limit (RL) for each chemical, yielding estimated concentrations above the RL but below all other measured concentrations (non ‘E’ values) for the associated chemicals. Occasionally, quality assurance issues at the NQWL resulted in exclusion of chemicals from the list of evaluated analytes. This resulted in slight differences in the numbers of chemicals monitored across site-years, with 176 chemicals monitored per site in 2017, and 165 – 172 chemicals monitored at per site in 2018. Detailed chemistry data including chemicals monitored, detection characteristics, RLs, and sample collection data is available in Table S3.

2.4.3. Evaluation of Chemical Data

Spatial and temporal trends in chemical concentrations across study sites were evaluated using multiple approaches. First, chemicals at the highest relative concentrations in study sites were identified by dividing molar concentrations of individual compounds by the cumulative molar concentration of detected mixtures within each site year. Chemicals contributing to ≥ 5 % of cumulative concentrations were identified as ‘important contributors’ to overall mixture concentration with each site-year. Second, sites were clustered based on relative concentrations of detected chemicals using hierarchical cluster analysis. To account for inter-year variability in study design and allow for a visual comparison of temporal trends, independent analyses were carried out for 2017 and 2018. Data were prepared by mean centering chemical concentrations and scaling to unit variance. Hierarchical cluster analysis was carried out using the ‘pvclust’ R (R Core Team 2020) package (Suzuki and Shimodaira 2006), with method.dist (distance metric) = ‘correlation’, method.hclust (linkage) = ‘average’, nboot (the number of bootstraps used for resampling) = 10000, and alpha = 0.80). Definition of site clusters was based on approximately unbiased (au) p-values, with final clusters being defined as those containing ≥ 2 sites and au > 0.80 (corresponding to p = 0.20). Chemical concentration profiles and cluster outputs were visualized using heat maps generated with the ‘ComplexHeatmap’ R package (Gu et al. 2016).

2.5. Whole Mixture Effect Assessment

2.5.1. in vitro Effect Assessment

Bioassays using a cell line stably transfected with human estrogen receptor alpha-regulated luciferase reporter gene constructs (T47D-KBluc; Wilson et al. 2004), were employed to estimate the total estrogenic activity of collected water samples (see Supplemental Information S1.1.1 for details). Briefly, estrogenic activity was quantified by interpolating the background-adjusted percent maximum response of each sample replicate against 17α-ethinylestradiol (EE2) standard curve concentrations using nonlinear regression (log agonist concentration vs. response-variable intensity; implemented in GraphPad Prism). Generated EE2-equivalent (EE2-EQ) concentrations (ng/L) were converted to estrogenic potency based on 17-β estradiol (E2-EQ; ng/L) by multiplying EE2-EQ by 0.68. Resulting E2-EQ were adjusted for sample dilution in the assay and averaged across replicates (n = 3), generating a mean concentration for each study site. Samples were considered significantly estrogenic if their activities exceeded three standard deviations above the mean assay control.

Non-targeted, in vitro high throughput screening of collected water samples was conducted through contract with Attagene Inc. (Morrisville, NC, USA) (see Supplemental Information S1.1.2 for details). Briefly, HepG2 cells were used to evaluate the impacts of water samples on transcription factor (cis-Factorial; 2017 – 2018), transfected nuclear receptor (trans-Factorial; 2017 – 2018), and G-protein coupled-receptor (GPCR-Factorial; 2018 only) activities representing a range of biological pathways (Romanov et al. 2008). Overall, 111 target endpoints were evaluated, including several related to xenobiotic and lipid metabolism, endocrine signaling, and cellular stress. Bioassay responses for individual filtered water samples (n = 2 per site) were expressed as fold-change relative to solvent controls (dimethyl sulfoxide), and response was batch-normalized based on response in the extraction blank run in each assay set. Furthermore, to account for any influences that sample collection, shipping, and testing methods may have had on measured in vitro activity, responses were normalized based on response observed in field blanks. Endpoints displaying activities ≥ 1.5-fold higher than the field blank responses were identified as ‘active’ (Blackwell et al. 2019).

2.5.2. in vivo Effect Assessment

To identify potential endocrine-related effects, in vivo measurements of the hypothalamus-pituitary-gonadal (HPG) axis status in caged fish were carried out by characterizing steroid hormone concentrations (17β-estradiol (E2) and testosterone (T)) in male and female fish plasma (2017) and vitellogenin (vtg) transcript abundance in hepatic tissues of male fish (2017 – 2018). To identify potential sublethal impacts of contaminants and/or contaminant mixtures on caged fish, abundance of various mRNA transcripts associated with xenobiotic metabolism were measured in intestinal (2017) and hepatic tissues (2017 and/or 2018) of exposed males. In 2017, measured transcripts included two cytochrome P450s (cyp 1A1, cyp 3A) and a UDP glucuronosyltransferase (ugt 1A1) in the hepatic and intestinal tissue and two cytochrome P450s (cyp 2D6, and cyp 2N13) in the intestinal tissue alone. In 2018, only cytochrome P450 (cyp 1A1) transcripts were measured in the hepatic tissue. Concentrations of E2 and T in caged fish plasma (ng/mL) were characterized by radioimmunoassays optimized for small volumes (Jensen et al. 2001). Abundances of mRNA transcripts in hepatic and intestinal tissues were evaluated via real-time quantitative, reverse-transcription polymerase chain reaction (qPCR) using kits, primers, and probe sequences detailed elsewhere (Cavallin et al. 2014, 2020; see Supplemental Information S1.1.3 for details). To accommodate the large number of samples required for transcript analysis (≤ 156 samples per tissue/transcript type) qPCR analyses were conducted across multiple assays. Thus, gene expression data were plate-normalized using response data from a common set of control samples run on each plate, correcting for potential variability across multiple plate runs (e.g., qPCR efficiency).

2.5.3. Analysis of Effect Data

Significant in vivo effects were identified by comparing measured responses of caged fish at study sites to those measured in reference water (MED). When residuals met normality (Shapiro-Wilk, p ≥ 0.05) and equality of variance (Levene’s test, p ≥ 0.05) assumptions, analysis of variance (ANOVA) followed by Dunnett’s post-hoc tests with Bonferroni corrections were applied for multiple comparisons analyses. Normality assumptions were evaluated using the ‘shapiro.test’ function and equality of variance was evaluated using the ‘leveneTest’ function, both available in the ‘stats’ R package (Chambers et al. 1992). One-way ANOVAs were carried out using the ‘aov’ function in the ‘stats’ R package and Dunnett’s post-hoc tests were carried out using the ‘DunnettTest’ function in the ‘DescTools’ R package (Signorell et al. 2021). When the data did not adequately adhere to normality and variance requirements, they were either log₁₀ transformed or analyzed using non-parametric Kruskal-Wallis tests followed by Dunn’s post-hoc tests with Bonferroni corrections. Kruskal-Wallis tests were carried out using the ‘kruskalTest’ function and Dunn’s post-hoc tests were carried out using the ‘kwManyOneDunnTest’ function in the ‘PMCMRplus’ R package (Pohlert 2021). Differences among sites were considered significant at p < 0.05. Due to sex-specific differences in physiology, in vivo effects were assessed independently for male and female fish (as applicable).

Site-specific trends in measured in vivo and in vitro effects were characterized using multiple approaches. First effect-based site groups were defined using hierarchical cluster analysis. Subsequently, trends in measured effect occurrence within site groups were characterized using principal components analysis (PCA). Effects considered in analyses were limited to those with at least one site that was determined to be significantly different from controls during pairwise testing (p < 0.05) or effect evaluation. Due to differences in the types of effects measured in each study year, 2017 and 2018 effect data were evaluated independently. Furthermore, one site (KKL) was excluded from 2017 analyses due to significant fish mortality (Section 3.1). Prior to all analyses, effect data were averaged, mean-centred, and scaled to unit variance. Hierarchical cluster analyses were carried out using the ‘pvclust’ R package (method.dist = ‘correlation’, method.hclust = ‘average’, nboot = 10000, and alpha = 0.80). Final clusters were selected as those containing ≥ 2 sites and displaying au ≥ 0.80. Relative responses across sites and final cluster groups were visualized using heat maps generated with the ‘ComplexHeatmap’ R package. PCAs were carried out using the ‘prcomp’ function in the ‘stats’ R package (R Core Team 2020), and final biplots were created using an correlation matrix. The first two principal component axes (PCA 1 + PCA 2) explained significant amounts of the overall variance in the 2017 and 2018 analyses (2017 PCA 1+PCA 2 variance = 69.7 %; 2018 PCA 1+PCA 2 variance = 80.8 %; Figure S2). Thus, important ecotoxicological effects in site groups were evaluated by assessing the direction and magnitude of effect loadings and the spatial distribution of sites on PCA 1 and PCA 2.

2.6. Component-Based Mixture Analysis

Component-based mixture analyses were carried out using concentration addition (CA) as a baseline concept for cumulative effects assessment. The CA concept assumes that cumulative toxicity reflects the sum of constituent concentrations (Loewe and Muischnek 1926). Thus, component-based analyses were carried out under the a priori assumption that chemicals detected at Milwaukee Estuary sites generally behaved in an additive manner when present in mixtures. This approach was primarily employed to simplify analyses, allowing for the prioritization of individual chemicals, mixtures, and study sites based on estimated cumulative effects and/or activities of complex and variable mixtures. However, to address the fact that similarly acting chemicals are more likely to act via CA, mixture groups were first classified based on similarity and cumulative effects were assessed under three distinct lines of evidence (LoE). First, significant drivers of apical effects were identified by comparing constituent concentrations to 96-h median lethal concentrations (LC50s) for fathead minnows (‘Apical LoE’). Second, significant drivers of sublethal effects were evaluated by comparing constituent concentrations to effect concentrations observed in in vitro assays (‘Non-Apical LoE’). Finally, significant drivers of pharmacological effects were evaluated by comparing constituent concentrations to measures of relative pharmaceutical potency (‘Pharmacological LoE’). Altogether, these LoE were chosen to evaluate potential mixture effects at diverse biological levels ranging from whole organism (‘Apical LoE’) to cellular/molecular (‘Non-Apical LoE’ and ‘Pharmacological LoE’). Within each LoE, mixture effects were assessed at three levels of increasing complexity: 1) individual constituent (single chemical); 2) mixture (simple mixture subsets); and 3) site-years (whole mixtures detected in each site and study year).

For each LoE, important predictors and mixtures were evaluated using MCRs (Price and Han 2011) calculated for each site-year in the Milwaukee Estuary. First, TQ were generated for chemical ($i$) at each site-year ($j$), by dividing the chemical concentration ($c_{i,j}$, μg/L) by an effect benchmark (${bench}_i$) (Eqn. 1):

$${TQ}_{i,j}=\frac{c_{i,j}}{{bench}_i}$$ (1)

Subsequently, MCRs were calculated for individual compounds (${MCR}_{i,j}$) by dividing the cumulative TQ at a site ($\Sigma {TQ}_j$) by the ${TQ}_{i,j}$ for an individual chemical (Eqn. 2):

$${MCR}_{i,j}=\frac{\sum {TQ}_j}{{TQ}_{i,j}}$$ (2)

Finally, MCRs were calculated for mixtures (${\mathrm{M}\mathrm{C}\mathrm{R}}_{\mathrm{m}\mathrm{i}\mathrm{x},\mathrm{j}}$) by dividing the cumulative TQ at a site-year ($\Sigma {TQ}_j$) by the cumulative TQ for a mixture (${TQ}_{mix,j}$) (Eqn. 3):

$${MCR}_{mix,j}=\frac{\sum {TQ}_j}{{TQ}_{mix,j}}$$ (3)

Chemicals or mixtures with MCRs ≤ 5 (contributing to ≥ 20 % of cumulative toxicity) were considered important contributors to cumulative effect at each site-year. For each LoE, important chemicals/mixtures were identified as those determined to be significant contributors to cumulative effect (MCR ≤ 5) in multiple site-years and/or those displaying a TQ > 0.01 and an MCR ≤ 5 in at least one site-year. Important site-years in the Milwaukee Estuary were identified by comparing TQ_j rankings across LoE. Constituent-based mixture analyses were carried out in R and visualized using the ‘ggpubr’ R package (Kassambara 2020). Methods employed for mixture group definition and benchmark derivation under each LoE are presented below.

2.6.1. Apical Effect Line of Evidence

Within the ‘Apical Effect’ LoE, mixtures were classified based on mode of action (MOA; e.g., narcosis (N), reactivity (S), or unknown/undefined (U)) and chemical structure (see Supplemental Information S1.1 for details). First, MOA was defined for each detected chemical based on the consensus of four different QSARs (Assessment Tools for Evaluation of Risk (Russom et al. 1991); TEST, v. 1.0.0 (Martin and Young 2001); the modified Verhaar classification scheme (Verhaar et al. 1992; Enoch et al. 2008), and the OASIS acute toxicity MOA tool (Mekenyan et al. 1990)), as described in Kienzler et al. (2019). Structural sub-groups were subsequently defined through hierarchical cluster analysis, using ToxPrint fingerprints (Chemotyper) (Yang et al. 2015) as structural identifiers. The ‘pvclust’ R package was used for hierarchical cluster analysis (method.hclust = complete, nboot = 10000, alpha = 0.70), applying the ‘dist.binary’ function in the ‘ade4’ package (Dray and Dufour 2007; Dray et al. 2021) to define distance matrices with the Hamman coefficient ((Gower and Legendre 1986); “method = 6”). Cluster analysis was performed independently for two MOA groups: 1) chemicals flagged as reactive (‘S’) and 2) chemicals flagged as narcotic (‘N’) in the consensus MOA evaluation. Furthermore, to account for uncertainties associated with the consensus MOA classifications, compounds unknown/undefined were grouped with both ‘N’ and ‘S’ chemicals during structural clustering. Final clusters were defined as those with 2 ≤ × < 10 chemicals and au ≥ 0.7 (Figure S3). Final mixture groups considered in apical LoE analyses were MOA-structure clusters containing ≥ 2 chemicals co-occurring in at least one site-year (Table S5).

‘Apical Effect’ benchmarks represented measured or estimated LC50s for fathead minnows (endpoint = mortality). Detailed methodology is available in the Supplemental Information (S1.2.2). Briefly, all available LC50s for the Milwaukee Estuary dataset were collected from the ECOTOXicology (ECOTOX) knowledgebase (U.S. EPA 2018). Search results were filtered to ensure toxicity tests were carried out with active ingredients of high chemical purity (≥ 70 %), and applied methods and reported values were accurate and statistically/toxicologically acceptable (e.g., adequate experimental design, appropriate control exposure, post-hatch life-stages, statistical significance of results). Chemicals with LC50s meeting filtering criteria were flagged as ‘data sufficient’ and geometric mean LC50s were used as apical effect benchmarks. For data-limited chemicals, LC50s were estimated via read-across or from the consensus of acute toxicity QSARs. Read-across was carried out using an extended chemical dataset representing all fathead minnow 96-h LC50s available in the ECOTOX Knowledgebase (Table S6). First, data were filtered to ensure study quality (as described above) and geometric means were computed, yielding one unique LC50 per chemical. Chemicals were grouped based on MOA and structure (as described above), and LC50 estimates were generated for data limited compounds using the weighted mean of log10-transformed LC50s (geomeans) from the ≤ 10 nearest chemical predictors within the cluster group. Nearest neighbours and weights used for read-across were based on distances between octanol/water partition coefficients; (Supplementary Information, S1.2.2). For the remaining (ungrouped) chemicals in the Milwaukee Estuary dataset, LC50 estimates were generated using consensus estimates from five acute toxicity QSARS: TEST (v. 5.1.1); Ecological Structure Activity Relationships class program (v 2.0; Mayo-Bean et al. 2012); and three VEGA HUB models (KNN/Read-Across fish acute (LC50) toxicity model v 1.0.0; NIC fish acute (LC50) toxicity model v 1.0.0; and the KNN/IRFMN Fathead Minnow LC50 model v 1.1.0) (Benfenati et al. 2013). The final list of 96-h LC50 values used as benchmarks under this LoE is available in Table S7.

2.6.2. Non-apical Effect Line of Evidence

Within the ‘Non-Apical Effect LoE’, mixtures were classified based on molecular targets and effect directions for active chemical-assay combinations available in the ToxCast in vitro database (United States Environmental Protection Agency 2015). Using the ‘toxEval’ R package (De Cicco et al. 2020), active assay endpoints for chemicals detected in the Milwaukee Estuary study sites were retrieved. Assay endpoints related to general physiology (e.g., “Cell Cycle”, “Background Measurement”, and “Cell Morphology”) and/or derived from platforms with largely nonspecific endpoints (e.g., Bioseek and Apredica), along with results flagged as “Borderline Active”, “Only highest conc above baseline active”, “Gain AC50 < lowest conc & loss AC50 < mean conc”, and “Biochemical assay with < 50% efficacy” were excluded (Kavlock et al. 2012; Blackwell et al. 2017). An additional 67 assays/assay-chemical combinations were excluded due to poor dose-response performance (e.g., nonmonotonicity), and results for 11 pharmaceuticals with limited ToxCast data were replaced with their salt-derivatives (Table S8). Chemical-assay outputs were annotated with assay-specific molecular targets and generalized effect directions (‘activation’, ‘binding’, and ‘inhibition’). Chemicals were subsequently grouped based on common molecular target and direction of effect (e.g., ‘ahR_activation’). Final mixtures used in MCR analyses represented molecular target-effect direction chemical groups containing ≥ 2 co-occurring chemicals (Table S9).

‘Non-Apical Effect’ benchmarks represented activity concentrations at cut-off (ACC) for chemical-assay combinations, collated from the ToxCast database. Prior to use in MCR analyses, ACCs were filtered to ensure data quality (as described above). For individual chemicals, final toxicity quotients represented the sum of exposure activity ratios (EAR; concentration/ACC) for each compound across all chemical-assay combinations in the filtered dataset (Blackwell et al. 2017). For mixtures, final toxicity quotients represented the sum of EARs for each applicable chemical-assay combination within defined molecular target - effect direction groups.

2.6.3. Pharmacological Effect Line of Evidence

Within the ‘Pharmacological Effect LoE’, mixtures were classified based on the molecular target and direction of activity of pharmaceuticals detected at Milwaukee Estuary study sites. Putative molecular targets and generalized directions of pharmacological activity (i.e., activation, inhibition, unknown) for detected pharmaceuticals were collated from the DrugBank database (DrugBank Online, v 5.0) (Wishart et al. 2018). To ensure specificity of pharmaceutical grouping, only targets with evidence of pharmacological action (defined in the DrugBank database as: Chemical Profile > Targets > Pharmacological action = Yes) were considered for mixture classification. Cross-species extrapolation was carried out employing the Sequence Alignment to Predict Across Species Susceptibility (SeqAPASS; v. 5.1; https://seqapass.epa.gov/seqapass/) tool (LaLone et al. 2016) to identify which pharmaceutical targets were likely present in fish species. Using universal protein resource (UniProt) identifiers obtained from the DrugBank database, primary amino acid sequences (Level 1) and putative functional domains of pharmaceutical targets (Level 2) were compared to protein sequences of all fish species available in the NCBI protein database (NCBI Resource Coordinators 2016). Parameters applied for SeqAPASS evaluations are available in Table S10. Molecular targets were considered ‘present in fish’ if any fish species had functional domains with sequence similarities above susceptibility cut-offs in either Level 1 or Level 2 SeqAPASS analyses. Pharmaceutical targets not ‘present in fish’ were not considered during mixture definition. Using the DAVID Bioinformatics Resources Gene Functional Classification Tool (v 6.8; Huang et al. 2009a,b), pharmaceutical targets were clustered into functionally related groups based on shared functional annotations. Pharmaceutical targets were identified using UniProt accession numbers (DrugBank), and gene functional classification was carried out under the highest stringency (kappa similarity term overlap = 5; kappa similarity threshold = 0.5; multiple linkage threshold = 0.5) modified to reflect the mixture definition applied in mixture analyses (initial group membership (min) = 2; final group membership (min) = 2). Mixtures were defined based on chemicals in common functional gene clusters with shared pharmacological targets and directions. Final mixtures were defined as target-direction groups with ≥ 2 chemicals that co-occurred within at least one study site (Table S11).

‘Pharmacological Effect’ benchmarks represented human peak therapeutic plasma concentrations (C_max), collated from the Mammalian Pharmacokinetic Prioritization for Aquatic Species Targeting (MaPPFAST) database (Berninger et al. 2016;Table S12). Two chemicals with available data in the DrugBank database (camphor and menthol) lacked available C_max values in the MaPPFAST database and were excluded from evaluation. For one measurement that was actually a mixture (pseudoephedrine/ephedrine) the lowest available C_max value (ephedrine C_max) was used. Representing the plasma concentration at which a drug exerts its therapeutic effect, C_max was applied as benchmark under this LoE to estimate the cumulative activities of pharmaceutical mixtures. However, as these C_max were primarily derived using mammalian data, there are limitations associated with their application for non-mammalian species (Berninger et al. 2016). For example, there are likely species-specific differences in adsorption, distribution, metabolism, and elimination (ADME), molecular-target binding affinities, and physiology which could influence drug potency/activity. As such, output from this LoE was considered with some caution, and high priority chemicals/mixtures were flagged as targets for further evaluation rather than definitive drivers of adverse ecotoxicological effects.

2.7. Effect Prediction

To bridge the gap between holistic and constituent-based mixture analyses, random forest regression paired with recursive feature elimination was used to identify chemicals and mixtures that were important predictors of measured effects. Detailed descriptions of random forest algorithms and their application in ecological classification and regression are available elsewhere (Vitense et al. 2019). This technique has been implemented for prioritization of single compounds detected in the Milwaukee Estuary (2017 – 2018), using the chemical and effect data described in this study. As such, detailed descriptions of applied techniques for the predictive analysis of this data are available elsewhere (companion paper). Briefly, random forest regression was carried out with the ‘cforest’ function in the ‘party’ R package with ntree (number of trees) = 5000, mtry (# predictors sampled at each node) = ceiling (# predictors/3), and other parameters set to those suggested in Strobl et al. (2007) for constructing unbiased random forests. Responses were defined as individual effects determined to be significantly different from control and/or blank responses in pairwise analyses (Tables S4 and S14). Predictors were defined as detected chemicals and all possible mixtures identified under the three LoEs (Sections 2.6.1 – 2.6.3), with individual chemical concentrations or cumulative mixture concentrations at each site-year being used as predictor inputs. Conditional permutation importance was carried out using the ‘permimp’ function in the ‘permimp’ R package (Debeer and Strobl, 2020), which quantifies predictor importance conditional on other correlated predictors in the model (conditional permutation threshold = 0.9). The least conditionally important predictors were removed iteratively, and importance was recalculated at each iteration (Gregorutti et al. 2016). The final model represented the subset of predictors that minimized the mean absolute “out of bag” error (i.e., prediction error obtained by predicting each data point using trees for which the data point was not in the bootstrapped training sample). Relationships between chemical/mixture concentrations and measured effects in the final model were evaluated by examining partial dependence plots. Important chemicals and/or mixtures were identified by separating measured effects into endocrine-related and xenobiotic response-related effect-categories (Table S4), and comparing selected predictor lists across different models. Important predictors were identified as those selected in > 20 % of effect models within each category.

3. Results and Discussion

3.1. Overview of Chemical Detection and Caged Fish Deployments in the Milwaukee Estuary

Composite water samples were successfully obtained from all 2017 study sites and most 2018 study sites, except for MIE and JIP, for which grab samples were collected upon cage retrieval (Table S3). Analytical characterization of collected water samples revealed that 77 unique chemicals were detected in the Milwaukee Estuary across study years (Table S3). In 2017, 22 % of monitored chemicals (39/176) were detected across study sites, including 16 pharmaceuticals and personal care products (PPCPs), 4 pesticides, 5 fuels and polycyclic aromatic hydrocarbons (PAHs), 2 wastewater indicators (WWIs), 9 industrial chemicals, and 3 fire retardants (Figure 1A, Table S3). In 2018, 42 % of monitored chemicals (70/165 – 168, depending on study site) were detected, including 35 PPCPs, 7 pesticides, 7 fuels or PAHs, 4 WWIs, 13 industrial chemicals, and 4 fire retardants (Figure 1B, Table S2). Caged fish deployments were largely successful across study years. Fish were shipped, received and deployed without incident. There were infrequent moralities, physical injuries, and/or deformities in some cages (survival = 83 – 100 % per cage; incident of injury/deformity = 0 – 20 % per cage; Table S4; Figure S6). However, comparative analyses demonstrated that occurrences of these apical effects were not significantly elevated compared to fish exposed to reference water (Table S14). Occasionally, fish were missing from cages (potentially due to predation, random mortality, or escape). When this occurred, final sample sizes were adjusted prior to use in mortality assessments. Finally, in 2017 there was mass mortality in KKL (survival = 0 % across all replicate cages; Table S4) likely due to dissolved oxygen deficiency in the river (average DO = 2.17 ± 0.88 mg/L Table S3). Thus, this site-year was excluded from all whole mixture assessments and predictive analyses using in vivo effect data.

Figure 1.

Spatial and temporal trends in concentrations of chemicals detected in grab or composite water samples collected from (A) 2017 and (B) 2018 study sites in the Milwaukee Estuary. Detection trends for individual chemicals across sites are visualized using hierarchical cluster analysis. Final site clusters are presented as column annotations (upper x-axes) and chemical classifications are presented as row annotations (left y-axes). Concentrations are mean-centred and scaled to unit variance on a chemical-basis, and scaled concentrations are presented in heatmaps with white boxes representing the lowest relative concentrations and dark blue boxes representing the highest relative concentrations. ‘MED’ refers to fish exposed to reference water in laboratory experiments carried out at the Great Lakes Ecotoxicology Division laboratory (Duluth, MN).

3.2. Whole Mixture Evaluation: Priority Sites in the Milwaukee Estuary

3.2.1. Spatiotemporal Patterns in Mixture Composition across Study Sites

Spatiotemporal analyses of analytical chemistry data highlighted trends that could be applied to characterize Milwaukee Estuary sites based on mixture composition (see details in the Supplemental Information S2.1). For example, three sites were grouped into site-clusters defined by relatively high concentrations of industrial compounds, PAHs, fuels, WWIs, and/or flame retardants in 2017 (MEC, MET) and/or 2018 (MEC, MET, MIN) (Figure 1; Table S13). Thus, these sites were characterized as being dominated by complex industrial/urban chemical mixtures (Table 1). Four sites were grouped into site-clusters defined by moderate-high concentrations of PPCPs in 2017 (MIE, MIP, MIM) and/or 2018 (MIE, MIP, MIM, CCM) (Figure 1; Table S13). Furthermore, evaluation of the heatmap demonstrated that JIP (which was not formally clustered with the other PPCP-dominated sites) also displayed relatively high PPCP concentrations (Figure 1B). Thus, these five sites were characterized as being primarily dominated by PPCP mixtures (Table 1). One site (MEF) displayed few detected chemicals at relatively low concentrations across both study years (Figure 1) and was characterized a relatively low impact site (Table 1). Finally, two sites (KKL and UCJ) displayed variable mixture compositions across study years. KKL displayed relatively high concentrations of select flame retardants, industrial compounds, pesticides, PPCPs, and industrial compounds in 2017, but relatively high concentrations of industrial compounds, PAHs, fuels, WWIs, and/or flame retardants in 2018 (Figure 1). Inversely, UCJ displayed high concentrations of industrial compounds, PAHs, fuels in 2017, and low concentrations of all measured contaminants in 2018 (Figure 1). As such, KKL and UCJ were characterized target sites for further monitoring, allowing for a more definitive profiles of mixture composition to be generated.

Table 1.

Equations used to identify priority chemicals and simplified mixture groups within component-based analyses of mixture effects. Definition of priority compounds and simplified mixture groups was based on the relative contribution of individual chemicals and simplified mixture groups to overall estimated ecotoxicity, bioactivity, or pharmacological effect at each individual site-year (defined as a cumulative ratio; CR), using analyses based on the maximum cumulative ratio (MCR) approach (Price and Han 2011).^*

Equation Number	Equation	Parameters	Explanation
Equation 1	${TU}_{i,j}=\frac{c_{i,j}}{{bench}_i}$	$i$ = chemical $j$ = site-year $c$ = concentration $bench$ = apical, non-apical, or pharmacological benchmark ${TU}_{i,j}$ = toxic units of chemical $i$ at site-year $j$	Toxic units were derived to estimate the relative effects of individual chemicals by relating detected concentration to a benchmark that describes apical, non-apical, or pharmacological effect in a target organism.
Equation 2	${\sum TU}_j={TU}_{i,1}+{TU}_{i,2}\dots +{TU}_{i,n}$	$j$ = site-year i1-n = chemicals detected at each site-year (1 – n) ${\sum TU}_j$ = cumulative toxic units at each site-year	Cumulative toxic units at each site-year were calculated by summing up the relative toxic units of each individual detected chemical.
Equation 3	${CR}_{i,j}=\frac{\sum {TU}_j}{{TU}_{i,j}}$	I = chemical $j$ = site-year $\Sigma {TU}_j$ = cumulative toxic units of site-year $j$ ${TU}_{i,j}$ = toxic units of chemical I at site-year $j$ ${CR}_{i,j}$ = cumulative ratio of chemical $i$ at site-year $j$	The cumulative ratio each chemical (${CR}_{i,j}$) was derived to describe how individual chemicals contribute to overall estimated apical, non-apical, or pharmacological effects at each site-year.
Equation 4	${\sum TU}_{mix,j}={TU}_{mixi,1}+{TU}_{mix,i2}\dots +{TU}_{mix,in}$	$mix$ = mixture group $j$ = site-year ${TU}_{mix,i,1-n}$ = toxic units for each constituent (i) in a mixture group ${\sum TU}_{mix,j}$ = cumulative toxic units for each mixture group at each site-year	Cumulative toxic units for each simplified mixture group at each site-year were calculated by summing up the relative toxic units of each individual detected chemical within the selected mixture group.
Equation 5	${CR}_{mix,j}=\frac{\sum {TU}_j}{{TU}_{mix,j}}$	${\sum TU}_{mix,j}$ = cumulative toxic units for each mixture group in each site-year $\sum {TU}_j$ = cumulative toxic units for each site-year ${CR}_{i,j}$ = cumulative ratio of a mixture group at site-year $j$	The cumulative ratio of mixture groups (${CR}_{mix,j}$) was derived to describe how simplified mixtures contribute to overall estimated apical, non-apical, or pharmacological effect at a site-year.

^*A detailed overview of component-based mixture analyses is available in Section 2.6.

3.2.2. Spatiotemporal Evaluation of Measured Effects at Study Sites

Spatiotemporal evaluation of measured effects highlighted trends that could be applied to characterize Milwaukee Estuary sites based on in vitro and in vivo responses to whole mixture exposures (see details in the Supplemental Information S2.2 and Tables S13 – S15). For example, four sites (MET, MEC, UCJ, and KKL) generally displayed xenobiotic response-related effects across evaluated years. In 2017, MEC, MET, and UCJ were clustered due to elevated responses related to xenobiotic metabolism (e.g., intestinal and hepatic cyp 1a1 and cyp 3a transcript expression, and elevated aryl hydrocarbon receptor (AhR) and pregnane-X-receptor (PXR) bioactivities), inflammation (e.g., elevated nuclear factor (erythroid-derived 2)-like 2 (NRF2) receptor (NRF2/ARE) bioactivity), and lipid homeostasis (e.g., elevated peroxisome proliferator response element (PPRE) and peroxisome proliferator receptor γ (PPARg) bioactivity) (Figures 2A and C; Table 1). In 2018 all four sites were clustered based on elevated bioactivities related to xenobiotic metabolism (e.g., AhR and PXR), inflammation (e.g., NRF2/ARE, prostaglandin receptor E2, D2, or I2 (PTGER2, PTGDR, PTGIR)), lipid homeostasis (e.g., PPRE, PPARg, peroxisome proliferator activated receptor α (PPARa), and retinoid-X-receptor β (RXRb)), cardiovascular activity (e.g., adrenoreceptor β1 (ADRB1) and/or oxidative stress (e.g., melanocortin 1 receptor (MC1R)) (Figures 2B and D; Table 1). Occasionally, these four sites displayed in vitro responses related to estrogen receptor activation (e.g., elevated estrogen receptor α (ERa), estrogen response element (ERE) bioactivity and/or elevated E2-EQ in collected water samples; Figures 2B and 2D; Table 1). However, elevated endocrine-related bioactivity was rarely observed in the same site across study years and no significant endocrine-related effects were observed in vivo. Thus, the weight-of-evidence indicated that whole mixtures occurring in these four sites primarily elicited xenobiotic response-related effects and they were flagged as high priority for xenobiotic response-related effect evaluation.

Figure 2.

Spatial and temporal trends in endocrine-related and xenobiotic response-related effects ^a measured in water and/or tissue samples collected from (A, C) 2017 and (B, D) 2018 caged fathead minnow studies in the Milwaukee Estuary watershed, visualized through (A – B) hierarchical cluster analysis and (C – D) principal components analysis (PCA). Visualized effects represent those that were significantly different from reference site fish (MED; p < 0.05), and effect data are mean-centred and scaled to unit variance on an effect-basis. In heat maps, final site clusters are presented as column annotations (upper x-axes), effect types are presented as row annotations (left y-axes), scaled effects are presented with a colour scale (white = lowest relative effect; dark blue = highest relative effect). The PCA plot presents the direction and magnitude of effect loadings and spatial distribution across sites on the first two principal component axes (PCA 1 + PCA 2)). ‘MED’ refers to fish exposed to reference water in laboratory experiments carried out at the Great Lakes Toxicology and Ecology Division laboratory (Duluth, MN).

^a Effect acronyms are defined as follows: E2_M_plasma = plasma 17β-estradiol concentrations (male fish); E2_F_plasma = plasma 17β-estradiol concentrations (female fish); AhR_bioactivity = aryl hydrocarbon receptor bioactivity; ADRB1_bioactivity = adrenoreceptor β1 bioactivity; ARE_ bioactivity = Antioxidant Response Element (ARE)-binding Nuclear factor (erythroid-derived 2)-like 2 (NRF2) bioactivity; CYP1A1_intestine = intestinal abundance of cytochrome P450 (CYP) 1A1 (male fish); CYP1A1_liver = hepatic abundance of CYP 1A1 (male fish); CYP2AD6_intestine = intestinal abundance of CYP 2AD6 (male fish); CYP2N13_intestine = intestinal abundance of CYP 2N13 (male fish); CYP3A_intestine = intestinal abundance of CYP 3A (male fish); CYP3A_liver = hepatic abundance of CYP 3A (male fish); E2-EQ = 17β-estradiol equivalents in grab/composite water; ERa_bioactivity= estrogen receptor α activity in in vitro assays; ERE_ bioactivity = estrogen response element activity in in vitro assays; GR_bioactivity = glucocorticoid receptor bioactivity; MC1R_bioactivity = melanocortin 1 receptor bioactivity; PPRE_bioactivity = peroxisome proliferator activating receptor bioactivity; PPARa_bioactivity = Peroxisome proliferator-activated receptor-a bioactivity; PPARg = peroxisome proliferator-activated receptor-γ bioactivity; PTGDR_bioactivity = prostaglandin D2 receptor bioactivity; PTGER2 = prostaglandin E receptor 2 activity in in vitro assays; RXRb_bioactivity = retinoid X receptor β bioactivity; PXRcis_bioactivity = pregnane-X-receptor bioactivity in cis-Factorial Attagene assays; PXRtrans_bioactivity = pregnane-X-receptor bioactivity in trans-Factorial Attagene assays; T_M_plasma = plasma testosterone concentrations (male fish); UGT1A1_liver = hepatic abundance of UDP glucuronosyltransferase 1A1 (male fish); UGT1A1_intestine = intestinal abundance of UDP glucuronosyltransferase 1A1.

Four other sites (MIM, MIP, MIE, JIP) could also be generally grouped together based on measured bioactivity and in vivo responses. In 2017, MIE, MIM, and MIP were clustered together (alongside MED and MEF; Table S13) due to elevated abundances of select xenobiotic metabolism related transcripts in hepatic and/or intestinal tissues (e.g., cyp 2n13, cyp 2ad6, ugt 1a1) and elevated endocrine-related responses (e.g., elevated E2 or decreased T concentrations in male fish, increased E2-EQ in collected water samples, and/or elevated ERE bioactivity; Figure 2A and 2C; Table 1). In 2018, MIE and MIM were grouped into a site-cluster, whereas JIP and MIP remained unclustered (Figure 2B; Table S13). However, evaluation of heatmaps and PCA biplots demonstrated that MIE and MIP displayed limited elevated responses whereas MIM and JIP displayed similar response profiles defined by elevated xenobiotic metabolism in hepatic tissues (cyp 1a1), endocrine-related bioactivity (E2-EQ and ERE), and inflammation (GR) (Figure 2B; Table 1). Altogether these four sites displayed slightly differing responses depending on the study year and/or suite of measured effects, but generally tended to display effects related to xenobiotic metabolism, endocrine pathways and/or inflammation. As such, these sites were flagged as high priority for xenobiotic response and endocrine related effect evaluation.

The three remaining sites (MEF, CCM, and MIN) displayed limited responses across evaluated years, demonstrating whole mixture effects similar to those observed in the laboratory controls (MED; Figure 2). For example, in 2017 MEF displayed some elevated bioactivity related to inflammation (NRF2/ARE) but relatively low expression of all other measured effects (Figures 2A and C; Tables 1 and S13). In 2018, CCM and MEF displayed moderately elevated bioactivities related to xenobiotic metabolism (PXR) and cardiovascular response (ADRB1) but low expressions of all other measured effects (Figures 2B and D; Table 1). Similarly, in 2018 MIN had low expressions of all evaluated responses (Figures 2B and D; Table 1). Overall, the weight-of-evidence indicated that (based on the analyses carried out in this study) whole mixtures occurring in MEF, CCM, and MIN did not significantly elevate in vivo or in vitro effects. Thus, these three sites were characterized as low priority for further effects assessments.

3.2.3. Site-Ranking based on Maximum Cumulative Ratio Analyses

Evaluation of whole mixtures using MCR analyses yielded relative rankings of estimated cumulative toxicity/activity for each site-year and across multiple LoE (Table 1; Tables S16 – S18). This allowed for study sites to be characterized based on their likelihoods to elicit apical, non-apical, and pharmacological effects. First, across evaluated year(s) MEC, MET, MIN, and UCJ displayed moderate-high cumulative apical toxicities (TQ = 0.007 – 0.03), variable non-apical activities (TQ = 0.01 – 0.23), and low-moderate pharmacological activities (TQ = <0.001 – 0.03) (Table 1; Figures 3–5). Similarly, KKL displayed high non-apical activities (TQ = 0.14 – 0.15) and low-moderate pharmacological activities (TQ = 0.002 – 0.03), but variable apical toxicity across study years (TQ = 0.003 (2017) – 0.06 (2018)) (Figures 3 – 5; Table 1). As such, these six sites demonstrated relatively high potentials for exerting adverse biological effects and were flagged as priorities for whole mixture assessments. Second, CCM, JIP, MIE, MIP, and MIM displayed low-moderate cumulative apical toxicities (TQ = 0.003 – 0.01), moderate-high cumulative non-apical activities (TQ = 0.02 – 0.16) and moderate-high cumulative pharmacological activities (TQ = 0.04 – 0.14) (Figures 3 – 5; Table 1). Thus, these five sites displayed the highest potentials for sublethal effects associated with pharmaceutical exposure and were flagged as priority sites for pharmaceutical mixture assessment. Finally, across all evaluated years MEF displayed relatively low cumulative toxicities/activities under apical (TQ ≤ 0.001), non-apical (TQ = <0.001 – 0.01), and pharmacological (TQ = 0) LoE (Figures 3 – 5; Table 1). Thus, this site displayed the lowest relative potential for adverse biological effects and was flagged as low priority for further mixture assessment.

Figure 3.

Comparison of the cumulative ratios (CR) of simplified mixture groups (A, C, E) and individual detected chemicals (B, D, F) to toxicity of whole mixtures detected at Milwaukee Estuary sites (2017 – 2018) using apical effect benchmarks (96-h lethal effect concentrations (LC50s), fathead minnows) to estimate cumulative effects. Toxic units (TU) of mixtures and chemicals of each site are compared across 2017 (A – B) and 2018 (C – D) to identify priority sites within each year. Important chemicals and mixtures are flagged as those that displayed CR exceeding 5 (equivalent to contributing to greater than 20 % of overall mixture pressure) across multiple site-years (A – D), or those that displayed CR > 5 and TU > 10⁻³ in at least one site-year (E – F).

*Grouping and components-based analyses were carried out using effect data collated from the ECOTOX Knowledgebase (United States Environmental Protection Agency, 2018) or estimated using read-across. Simplified mixture groups were defined based on putative mode of action (Kienzler et al., 2019) and structural similarity, and CR analyses were carried out by comparing chemical concentrations at Milwaukee Estuary study sites to 96-h lethal effect concentrations (LC50s) for fathead minnows measured or estimated for detected chemicals.

Figure 5.

Comparison of the cumulative ratios (CR) of simplified mixture groups (A, C, E) and individual detected chemicals (B, D, F) to putative pharmacological activity of whole mixtures detected at Milwaukee Estuary sites (2017 – 2018) using pharmacological benchmarks (mammalian maximum serum concentrations (C_max), MaPPFAST database) to estimate cumulative effects. Toxic units (TU) of mixtures and chemicals of each site are compared across 2017 (A – B) and 2018 (C – D) to identify priority sites within each year. Important chemicals and mixtures are flagged as those that displayed CR exceeding 5 (equivalent to contributing to greater than 20 % of overall mixture pressure) across multiple site-years (A – D), or those that displayed CR > 5 and TU > 10⁻³ in at least one site-year (E – F).

* Grouping and component-based analyses were carried out using mammalian pharmacological data derived from the DrugBank online database (Wishart et al., 2006) and functional annotations of mammalian genes (Huang et al., 2009a, 2009b). CR analyses were carried out by comparing measured chemical concentrations in the Milwaukee Estuary (2017 – 2018) to mammalian maximum serum concentrations (C_max) collated from the MaPPFAST database (Berninger et al., 2016).

3.2.4. Synthesis of Site-Specific Analyses and Identification of Target Sites

The first objective of this study was to employ and integrate evaluative multiple techniques to identify priority sites in the Milwaukee Estuary based on whole mixture toxicity. As such, spatiotemporal trends in mixture composition and in situ effects were considered alongside estimated apical, non-apical, and pharmacological toxicities/activities to bin and prioritize sites under distinct action categories (Table 1).

Amongst sites surveyed, KKL, MEC, MET, and UCJ demonstrated the greatest potentials to be adversely impacted by complex industrial/urban mixtures (Table 1). In at least one study year these sites: 1) were dominated by mixtures of PAHs, fuels, industrial chemicals, WWIs, and flame retardants; 2) demonstrated altered xenobiotic and lipid metabolism, inflammation, estrogen receptor activation, cardiovascular activity, and/or oxidative stress response; and 3) displayed moderate-high potentials for apical toxicity and/or non-apical activity. Anthropogenic land-use trends indicated that these sites are highly urbanized, dominated by residential areas, commercial and industrial operations, roads, and parking lots (Table S1; (Falcone 2015)). Thus, it is likely that these complex mixtures originated from industrial discharge, combustion-related deposition, run-off from road systems, coal-tar sealed parking lots urban run-off, and/or sewer overflows (Yunker et al. 2002; Crane 2014; Abdel-Shafy and Mansour 2016; Baldwin et al. 2017). These complex urban/industrial mixtures are of potential ecotoxicological concern as (along with the sub-lethal, whole mixture effects observed in this study) their constituents have been shown to impact fish cardiac function and morphological development (Honda and Suzuki 2020), be photoactivated to more reactive forms under UV-light exposure (Ankley et al. 2003), accumulate in sediments until they reach concentrations toxic to benthic invertebrates (Baldwin et al. 2017), and/or disrupt endocrine systems in aquatic biota (Frenzilli et al. 2021; Rosenmai et al. 2021). Furthermore, some of these constituents have been shown to act additively (e.g., Erickson et al., 1999), which could exacerbate cumulative effects. As this study primarily focused on evaluating effects of short-term (96-h) exposures in fish, additional analyses are required to definitively characterize the chronic and/or ecosystem-level impact(s) of these whole mixtures. However, all available evidence indicates that MEC, MET, KKL, and UCJ display the highest potential for adverse effects associated with complex urban/industrial mixture exposure. Hence these four sites were flagged as high priority for additional hypothesis-driven monitoring of complex urban/industrial mixtures in the Milwaukee Estuary.

Compared to the other Milwaukee Estuary sites, JIP, MIM, MIP, and MIE demonstrated the greatest potential to be adversely impacted by PPCP mixtures (Table 1). These sites tended to display: 1) relatively high concentrations of PPCPs across evaluated years; 2) elevated effects associated with xenobiotic metabolism, inflammation, estrogen receptor activation and/or in vivo hormone alteration; 3) and moderate-high potentials for pharmacological and/or non-apical activity. Anthropogenic land-use trends demonstrated that these sites were located in mixed land-use areas, displaying land-use profiles dominated by agriculture, low-use/conservation areas, and high-medium density residential areas (Table S1; (Falcone 2015)). Furthermore, two of these sites (JIP and MIM) were closely located, downstream of the Jones Island Water Reclamation Facility. Therefore, it is likely that the pharmaceutical mixtures originated from WWTP discharges, urban run-off, or sewer overflows (Lajeunesse et al. 2012; Tran et al. 2018; Yang et al. 2019; Golovko et al. 2021). These PPCP mixtures are of potential ecotoxicological concern because many of the constituents are formulated to elicit specific pharmacological activities (e.g., inhibition or activation of cation channels, neurotransmitters, receptors, enzymes, transcription factors, or other functional proteins; Table S11). Consequently, it is possible that these compounds could act on similar molecular targets and elicit sub-lethal effects in aquatic biota. Indeed, along with the endocrine- and xenobiotic metabolism-related effects observed here, some of the PPCPs detected in the Milwaukee Estuary have been shown to adversely impact reproduction, growth, behaviour, development, and other physiologically important functions of aquatic biota in laboratory and field studies (Sanchez et al. 2011; Burkina et al. 2015; Niemuth et al. 2015). Furthermore, although the mixtures of PPCPs detected in this study have yet to be formally tested, mixture constituents could act additively (Godoy and Kummrow 2017) enhancing their effects. Overall, the available evidence indicates that JIP, MIM, MIE, and MIM display the highest potential for adverse effects associated with PPCP mixture exposure. Accordingly, these sites were flagged as high priority for additional hypothesis-driven monitoring of PPCP mixtures in the Milwaukee Estuary.

Synthesis of results from site-specific analyses also allowed for the identification of potential low priority sites in the Milwaukee Estuary. For example, across all study years, MEF had low chemical detection, limited elevated in vivo or in vitro effects, and low TQs across all LoE (Table 1). Furthermore, this site displays land-use profiles dominated by crops, pastures, and low-use areas (SI Table 1; (Falcone 2015)), indicating that high-concentration input of PAHs, fuels, PPCPs, or other industrial/mixed-use chemicals is unlikely. Consequently, MEF was flagged as low priority site for cumulative effect evaluation, and a candidate reference site for future effect-based mixture studies in the Milwaukee Estuary. Alternatively, MIN had elevated concentrations of industrial chemicals and WWI and moderate apical TQ but limited measured effects and low non-apical and pharmacological TQs (Table 1). Similarly, CCM had relatively high concentrations of PPCPs and moderate-high non-apical and pharmacological TQs but limited measured effects and low apical TQs (Table 1). Thus, although concentration-based analyses identified MIN and CCM as a potential high priority sites for mixture assessments, these two sites were flagged as low priority due to a general lack of concordance with measured responses (Table 1). Interestingly, although both sites were located downstream of WWTP, they displayed differing chemical detection and measured effect profiles than JIP and MIM. This could likely be attributed to differences in WWTP proximity and/or other land-use characteristics. For example, MIN was located approximately 5 miles downstream of the Village of Newburg Sanitary Sewer Treatment Facility in an area dominated by agriculture, low-use/conservation areas, and low-medium density residential areas (Table S1 (Falcone 2015). Thus, this site was likely subjected to low overall chemical input, yielding limited observed effects. On the other hand, CCM was around 1000 feet downstream of the Cedarburg Wastewater Plant, but displayed a land-use profile dominated by agriculture, low-use/conservation areas, and low-medium density residential areas (Table S1 (Falcone 2015)). As such, CCM likely had lower overall PPCP inputs than MIM and JIP (which were located in high-medium residential areas), perhaps limiting observed responses in the in situ studies.

3.3. Component-Based Evaluation: Priority Chemicals and Mixtures in Milwaukee Estuary

3.3.1. Maximum Cumulative Ratio Analysis under Diverse Lines of Evidence

Grouping of chemicals detected in the Milwaukee Estuary based on MOA-structure (apical LoE), in vitro targets and assay directions in the ToxCast database (non-apical LoE), and fish-susceptible targets and MOAs of pharmaceuticals (pharmacological LoE) yielded 144 unique mixture groups. Eighteen mixtures were defined under the apical LoE (Table S5), 120 mixture groups were defined under the non-apical LoE, (Table S9), and six mixture groups were defined under the pharmacological LoE (Table S11). Depending on the specific LoE under evaluation, between 13 and 21 chemicals remained ungrouped for MCR analyses.

Using measured/estimated 96-h fathead minnow LC50s as effect benchmarks, two mixtures and three chemicals were identified as important contributors to cumulative ecotoxicological effects (Figure 3, Tables 2 – 3). Amongst the mixtures, PAHs were the most significant contributors to cumulative toxicity, displaying MCRs ≤ 5 (indicating contribution to ≥ 20 % of cumulative toxicity; (Price and Han 2011)) in 17 site-years and TQs > 0.01 in 3 site-years (Figures 3A and C; Tables 2 and S15). Conversely, sterol mixtures displayed MCR ≤ 5 in 4 site-years but did not exceed TQ = 0.01 (Figures 3A and C; Tables 2 and S15). Amongst the individual chemicals, fluoranthene [CAS = 206–44-0] largely dominated ecotoxicological profiles, displaying MCRs ≤ 5 in 17 site-years and TQs > 0.01 in 3 site-years (Figures 3B and D; Tables 3 and S15). In contrast, β-sitosterol [CAS = 83–46-5] and cholesterol [CAS = 57–88-5]) displayed MCRs ≤ 5 in 4 and 2 site-years, respectively, but did not exceed TQ = 0.01 (Figure 3B, Tables 3 & S16). Overall, PAH mixtures and fluoranthene were flagged as high priority; β-sitosterol, cholesterol, and sterol mixtures were flagged as medium priority; and the remaining 74 individual chemicals and 16 mixtures were flagged as low priority under the apical LoE.

Table 2.

Definition and prioritization of Milwaukee Estuary study sites (2017 – 2018) through a holistic, integrated mixture assessment. Sites were divided into differing groups based on their chemical composition. Sites were subsequently prioritized based on endocrine-related (ER) and xenobiotic response related (XR) effects measured in situ studies and cumulative toxicity quotients (ΣTQ) or exposure activity ratios (EAR) derived in component-based analyses carried out using apical, non-apical, and pharmacological data.

Sites	Study Year	Chemical Composition			Measured Effects		ΣTQ/EAR			Final Site Prioritization
		Detected Chemicals (n)	Dominant Chemical Classes	Cluster Group	XR	ER	Apical Data	Non-Apical Data	Pharmacological Data
Sites Defined by Complex Mixtures of Industrial and Urban Contaminants
Menomonee River near Church St. at Wauwatosa (MEC)	2017	25	PAHs, Fuels, Industrial	Industrial/urban chemical mixtures	AhR/PXR-related metabolism, inflammation, lipid homeostasis	Estrogen receptor activation	0.01	0.23	0.03	High Priority for Non-Targeted Effects Assessment
	2018	25	OP-FRs, PAHs, Fuels, Industrial, Pesticides	Industrial/urban chemical mixtures	AhR/PXR-related metabolism,	-	0.02	0.04	0.03
					inflammation, oxidative stress response
Menomonee River at 25th Street (MET)	2017	17	PAHs, Fuels, Industrial	Industrial/urban chemical mixtures	PXR/AhR-related metabolism, inflammation, lipid homeostasis	Estrogen receptor activation	0.01	0.02	0.01	High Priority for Non-Targeted Effects Assessment
	2018	28	OP-FRs, PAHs, Fuels, Industrial, Pesticides, WWIs	Industrial/urban chemical mixtures	PXR/AhR-related metabolism, inflammation, cardiovascular response, oxidative stress response	-	0.03	0.06	0.03
Milwaukee River at County Hwy M near Newburg (MIN)	2018	25	Industrial, WWIs	Industrial/urban chemical mixtures	NA – low impact site.	-	0.01	0	0	Low Priority for Non-Targeted Effects Assessment
Sites Defined by Complex Mixtures of Pharmaceuticals and Personal Care Products
Milwaukee River at Estabrook (MIE)	2017	27	PPCPs	PPCPs	Intestinal/hepatic drug metabolism	Hormone alteration	0	0.16	0.14	High Priority for Targeted (Pharmaceutical) Effects Assessment
	2018	25	PPCPs, Pesticides	PPCPs	AhR-related metabolism	-	0.01	0.02	0.05
Milwaukee River at Mouth (MIM)	2017	16	PPCPs	PPCPs	Intestinal/hepatic drug metabolism	Hormone alteration	0	0.01	0.04	High Priority for Targeted (Pharmaceutical) Effects Assessment
	2018	30	PPCPs, Pesticides	PPCPs	AhR-related metabolism, inflammation	Estrogen receptor activation	0.01	0.02	0.06
Milwaukee River at Walnut Street (MIP)	2017	23	PPCPs	PPCPs	Intestinal/hepatic drug metabolism,	Hormone alteration	0	0.08	0.12	High Priority for Targeted (Pharmaceutical) Effects Assessment
					AhR/PXR-related metabolism, lipid homeostasis
	2018	25	PPCPs, Pesticides	PPCPs	PXR-related metabolism, cardiovascular response	-	0.01	0.02	0.06
Cedar Creek at Green Bay Road at Cedarburg (CCM)	2018	41	PPCPs	PPCPs	PXR-related	-	0	0.04	0.12	Low Priority for Targeted (Pharmaceutical) Effects Assessment
					metabolism, cardiovascular response
Ungrouped Due to Variable or Undefined Spatiotemporal Composition(s)
Kinnickinnic River at Lincoln (KKL)	2017	18	OP-FRs, Industrial, Pesticides, PPCPs	Ungrouped	<Not evaluated>	-	0	0.15	0	High Priority for Further Mixture Characterization
	2018	33	OP-FRs, Industrial, PAHs,	Industrial/urban chemical mixtures	AhR/PXR-related metabolism, inflammation, cardiovascular response, oxidative stress response	-	0.06	0.14	0.03
			Fuels, Pesticides, WWI
Underwood Creek at Juneau Blvd (UCJ)	2017	15	PAHs, Fuels, Industrial	Industrial/urban chemical mixtures	AhR/PXR-related metabolism, inflammation, lipid homeostasis	Estrogen receptor activation	0.01	0.01	0	High Priority for Further Mixture Characterization
	2018	14	NA – low relative detection.	Low impact	AhR/PXR-related metabolism, inflammation, oxidative stress response	-	0.01	0.01	0
Milwaukee Outer Harbor - Jones Island Plume (JIP)	2018	46	Industrial, PPCPs	Ungrouped	AhR-related metabolism, inflammation	Estrogen receptor activation	0.01	0.03	0.1	High Priority for Further Characterization of Mixture Composition
Low Impact Site(s)
Menomonee River at Freistadt Road (MEF)	2017	4	NA – low relative detection.	Low impact	Inflammation	-	0	0.01	0	Low Priority
	2018	3	NA – low relative detection.	Low impact	PXR-related metabolism, cardiovascular response	-	0	0	0

^aWWIs = wastewater indicators; PPCPs = pharmaceuticals and personal care products; Industrial = industrial and mixed-use chemicals; OP-FRs = organophosphate flame retardants.

^bAhR = Aryl hydrocarbon receptor; AhR metabolism = elevated expression of cytochrome P450 (cyp) 1A1 in intestinal or hepatic fish tissues or AhR bioactivity; Cardiovascular response = elevated adrenergic β1 receptor bioactivity; Estrogen receptor activation = elevated bioactivity of estrogen receptor α and/or estrogen response element, or elevated concentrations of 17β-estradiol equivalents in collected water samples; Hormone alteration = elevated concentrations of 17β-estradiol in male fish plasma; Inflammation = elevated bioactivity of antioxidant response element-binding nuclear factor like 2 or prostaglandin (D2, E2, or I2) receptors; Intestinal/hepatic drug metabolism = elevated expression of cyp 2ad6, cyp 2n13 and ugt 1a1 in intestinal and/or hepatic fish tissues; Lipid homeostasis = elevated peroxisome proliferator receptor (PPRE, PPARa, PPARg) or retinoid-X-receptor bioactivity; Melanin production = elevated melanocortin 1 receptor bioactivity; PXR-related metabolism = elevated cyp 3a expression in hepatic and intestinal fish tissues and/or PXR bioactivity.

Table 3.

Simplified mixture groups identified as important contributors to mixture effects in the Milwaukee Estuary (2017 – 2018). Important drivers of mixture effects were evaluated using apical, non-apical, and pharmacological effect data and cumulative ratio (CR) analyses. Mixture groups representing important predictors of endocrine-related (ER) and xenobiotic response-related (XR) effects measured in in situ caged fish studies were identified using random forest regression. High, medium, and low priority mixtures were identified based on the consensus of predictions generated from component-based and random forest analyses.

Mixture	Constituents	CR Analyses			Predictive Analyses		Mixture Prioritization
		Data Type	CR < 5 (n site-years)	TQ ≥ 10−2 (n site-years)	Data Type	Proportion of Selected Effects	Priority Group	Supporting Data Types
PAHs	Anthracene, Benzo[a]pyrene, Fluoranthene, Phenanthrene, Pyrene	Apical	17	3	-	<0.2	High Priority	Apical
ESR1_activation	3,4-dichlorophenyl isocyanate, 4-tert-octylphenol, Anthraquinone, Benzo[a]pyrene, Bisphenol A, Caffeine, Carbamazepine, Carbaryl, Fluoranthene, Metolachlor, Nicotine, Pentachlorophenol, Pyrene, Triamterene, Tributyl phosphate, Triclosan, Tris(2-butoxyethyl)phosphate	Non-apical	4	4	XR	0.25	High Priority	Non-Apical; XR Effect Prediction
SLC_inhibition	Bupropion, Citalopram, Desvenlafaxine, Pseudoephedrine + Ephedrine, Tramadol, Venlafaxine	Pharmacological	9	8	ER	0.4	High Priority	Pharmacological; ER Effect Prediction
NR1I2_activation	Atrazine, Bisphenol A, Carbamazepine, Carbazole, N,N-diethyltoluamide, Fluoranthene, Indole, Isophorone, Metolachlor, Phenanthrene, Pyrene, Tris(2-butoxyethyl)phosphate, Tris(2-chloroethyl)phosphate	Non-apical	2	0	XR	0.25	Medium Priority	Non-Apical; XR Effect Prediction
Sterols	β-sitosterol, Cholesterol	Apical	4	0	ER	0.2	Medium Priority	Apical; ER Effect Prediction
CYP2B6_activation	Atrazine, Metolachlor	Non-apical	3	0	-	<0.2	Medium Priority	Non-Apical
NFE2L2_activation	3,4-dichlorophenyl isocyanate, Benzo[a]pyrene, Caffeine, Metolachlor, Triamterene	Non-apical	1	1	-	<0.2	Medium Priority	Non-Apical
NR1I3_binding	Bisphenol A, Fluoranthene, Metolachlor	Non-apical	1	1	XR	0.25	Medium Priority	Non-Apical
TP53_activation	Benzo[a]pyrene, Methyl-1H-benzotriazole	Non-apical	2	0	-	<0.2	Medium Priority	Non-Apical
Industrial_PPCP_Narcotics	Anthraquinone, Camphor, Cotinine, Galaxolide, Isophorone, Isoquinolone, Nicotine	Effect Prediction	0	0	ER XR	0.4 0.25	Medium Priority	ER Effect Prediction; XR Effect Prediction
Androstenedione_inhibition	Bisphenol A, Fluoranthene, Metoprolol, Pyrene	Effect Prediction	0	0	XR	0.25	Medium Priority	XR Effect Prediction
AR_activation	4-tert-octylphenol, Benzo[a]pyrene, Bisphenol A, Bromoform, Carbamazepine, Fluoranthene, Methyl salicylate, Ranitidine, Tolytriazole, Triamterene, Triclosan, Tris(1,3-dichloro-2-propyl) phosphate	Effect Prediction	1	0	XR	0.25	Medium Priority	XR Effect Prediction
PGR_inhibition	3,4-dichlorophenyl isocyanate, 4-tert-octylphenol, Benzo[a]pyrene, Bisphenol A, Carbaryl, Carbazole, Fluoranthene, Methotrexate, Metolachlor, Pentachlorophenol, Piperonyl butoxide, Pyrene, Triclosan, Triphenyl phosphate, Tris(1,3-dichloro-2-propyl) phosphate, Tris(2-butoxyethyl)phosphate	Effect Prediction	0	0	XR	0.2	Medium Priority	XR Effect Prediction
RXRa_activation	4-tert-octylphenol, Benzo[a]pyrene, Diphenhydramine, Methyl salicylate, p-Cresol, Triamterene, Triclosan, Trimethoprim	Effect Prediction	0	0	XR	0.2	Medium Priority	XR Effect Prediction

Using ToxCast ACC as effect benchmarks, six mixtures and seven chemicals were identified as important contributors to cumulative bioactivity (Figure 4; Tables 2 – 3). Amongst the mixtures, ESR1_activation mixtures were the most significant contributors to cumulative activity, displaying MCRs ≤ 5 and TQ > 0.01 in 4 site-years (Figures 4A and C; Tables 2 and S17). Two additional mixtures (NFE2L2_activation and NR1I3_binding) displayed MCRs ≤ 5 and TQ > 0.01 in 1 site-year, and three mixtures (CYP2AB6_activation, NR1I2_activation, and TP53_activation) displayed MCRs ≤ 5 in 2 – 3 site-years but did not exceed TQ = 0.01 (Figures 4A and C; Tables 2 and S17). Amongst the individual chemicals, metolachlor [CAS = 51218–45-2], benzo[a]pyrene [CAS = 50–42-8], and bisphenol A [CAS = 80–05-7] were the most significant contributors to cumulative bioactivity, displaying MCRs ≤ 5 in 3 – 8 site-years and TQ > 0.01 in 1 – 3 site-years (Figures 4B and D, Tables 3 and S17). Conversely, 4-tert-octylphenol [CAS = 140–66-9], methotrexate [CAS = 59–05-2], tributyl phosphate [CAS = 126–73-8], and tris(2-butoxyethyl)phosphate [CAS = 78–51-3] only displayed MCRs ≤ 5 and TQ > 0.01 in 1 site-year each (Figures 4B and D; Tables 3 and S17). Overall, ESR1_activation mixtures, metolachlor, benzo[a]pyrene, and bisphenol A were flagged as high priority; 4-tert-octylphenol, methotrexate, tributyl phosphate, tris(2-butoxyethyl)phosphate, and NFE2L2_activation, NR1I3_binding, CYP2AB6_activation and NR1I2_activation, TP53_activation mixtures were flagged as medium priority; and the remaining 70 chemicals and 112 mixtures were designated as low priority under the non-apical LoE.

Figure 4.

Comparison of the cumulative ratios (CR) of simplified mixture groups (A, C, E) and individual detected chemicals (B, D, F) to the in vitro bioactivity of whole mixtures detected at Milwaukee Estuary sites (2017 – 2018) using non-apical effect benchmarks (exposure-activity ratios (EARs), ToxCast database) to estimate cumulative effects. Toxic units (TU) of mixtures and chemicals of each site are compared across 2017 (A – B) and 2018 (C – D) to identify priority sites within each year. Important chemicals and mixtures are flagged as those that displayed CR exceeding 5 (equivalent to contributing to greater than 20 % of overall mixture pressure) across multiple site-years (A – D), or those that displayed CR > 5 and TU > 10⁻³ in at least one site-year (E – F).

*Grouping and component-based analyses were carried out using in vitro effect data collated from the ToxCast database (Judson et al., 2009; United States Environmental Protection Agency, 2015), with simplified mixture groups being defined based on molecular target and effect direction of active in vitro assays. CR analyses were carried out by comparing chemical concentrations at Milwaukee Estuary study sites to activity concentrations at cut-off (ACC) for applicable chemical-assay combinations.

Using MaPPFAST Cmax as effect benchmarks, one mixture and three chemicals were identified as important contributors to cumulative pharmacological activity. SLC_inhibition mixtures largely dominated cumulative pharmacological profiles, displaying MCRs ≤ 5 in 9 site-years and TQ > 0.01 in 8 site-years (Figures 5A and C; Tables 2 and S18). Of the individual chemicals, desvenlafaxine [CAS = 93413–62-8] and nicotine [CAS = 54–11-5] were the most significant contributors to pharmacological effects, displaying MCRs ≤ 5 in 7 – 8 site years, and TQ > 0.01 in 6 – 8 site years (Figures 5B and C; Tables 3 and S18). Conversely, metformin [CAS = 657–24-9] displayed MCR ≤ 5 in 2 site-years and did not exceed TQ = 0.01 (Figures 5B and C; Tables 3 and S18). Overall, SLC_inhibition mixtures, desvenlafaxine, and nicotine were flagged as high priority; metformin was flagged as medium priority; and the remaining 34 pharmaceuticals and five mixtures were flagged as low priority under the pharmacological LoE.

3.3.2. Predictive Analyses

Random forest analyses identified a suite of chemicals and mixtures that were important predictors of in vitro and in vivo effects measured in caged fish studies (Table S19, Figures S11 – S16). Overall, five in vitro effects (ADRB1, MC1R, PTGIR, PTGDR, PTGER2) were excluded from predictive analyses due to model failure (Table S19). Weight-of-evidence evaluation of the model outputs yielded three mixtures and two single compounds that were important predictors of endocrine related effects (Tables 2 – 3; Figure 6). Two of these mixtures (Sterols, SLC_inhibition) and both individual compounds (β-sitosterol [CAS = 83–46-5] and desvenlafaxine [CAS = 93413–62-8]) were previously identified as high or medium priority under MCR LoE (Tables 2–3). The remaining mixture identified as an important predictor of endocrine-related effects was a complex mixture of industrial chemicals and PPCPs (industrial_PPCP_narcotics) (Table 2). Seven mixtures and four single compounds were determined to be important predictors of xenobiotic response related effects (Tables 2 – 3; Figure 6). Two of these mixtures (ESR1_activation and NR1I2_activation) and one of these individual compounds (metformin [CAS = 657–24-9]) were previously identified as high or medium priority under MCR LoE (Tables 2 – 3). The remaining mixtures included the industrial_PPCP_narcotics mixture, and four ToxCast mixtures (androstenedione_inhibition, AR_activation, PGR_inhibition, and RXRa_activation) (Table 2). The remaining three chemicals identified as important effect predictors were anthraquinone [CAS = 84–65-1] acetaminophen [CAS = 103–90-2] and tramadol [CAS = 27203–92-5]. Overall, predictive analyses combined with weight-of-evidence identified eight mixtures and six individual compounds that that could elicit endocrine- and/or xenobiotic metabolism-related effects, adding five mixtures and three compounds to the list of medium-high priority targets for ecotoxicological effects evaluation.

Figure 6.

Individual mixture constituents and mixture groups identified as important predictors of statistically significant in vivo and in vitro effects ^a in caged fish studies. Important predictors were identified through random forest regression paired with recursive feature elimination. Measured effects are grouped based on effect type (top axis; ER = endocrine related effect; XM = xenobiotic metabolism-related effect) and chemical mixture constituents/mixture groups are identified based on grouping strategy (right axis; MOA = grouped based on structure and mechanism of action; ToxCast = grouped based on active targets in ToxCast in vitro database; Pharm = grouped based on pharmacological targets and generalized direction of activity using mammalian pharmacological data). Important (selected) predictors are highlighted in black and unimportant (unselected) predictors are highlighted in yellow.

^a ahR = aryl hydrocarbon receptor activity in in vitro assays; ADRB1 = Adrenoreceptor β1 (ADRB1) activity in in vitro assays; ARE = Antioxidant Response Element (ARE)-binding Nuclear factor (erythroid-derived 2)-like 2 (NRF2) (NRF2/ARE) activity in in vitro assays; CYP1A1_intestine = intestinal abundance of cytochrome P450 (CYP) 1A1 (male fish); CYP1A1_liver = hepatic abundance of CYP 1A1 (male fish); CYP2AD6_intestine = intestinal abundance of CYP 2AD6 (male fish); CYP2N13_intestine = intestinal abundance of CYP 2N13 (male fish); CYP3A_liver = hepatic abundance of CYP 3A (male fish); CYP3A_intestine = intestinal abundance of CYP 3A (male fish); Era_trans = estrogen receptor alpha activity in in vitro assays; ERE_cis = estrogen response element activity in in vitro assays; GSI = gonadosomatic index; GR = Glucocorticoid Receptor (GR) activity in in vitro assays; PPARa = peroxisome proliferator-activated receptor-α activity in in vitro assays; PPARg_trans = peroxisome proliferator-activated receptor-γ activity in in vitro assays; PPRE = peroxisome proliferator activating receptor activity in in vitro assays; PXR_cis = pregnane-X-receptor activity in cis-factorial in vitro assays; PXR_trans = pregnane-X-receptor activity in trans-factorial in vitro assays; RXRb = retinoid-X-receptor β activity in in vitro assays; RIA_F = plasma estradiol concentrations (female fish); RIA_M = plasma estradiol concentrations (male fish); T47_D= estradiol equivalents in grab/composite water; UGT1A1_liver = hepatic abundance of UDP glucuronosyltransferase 1A1 (male fish); UGT1A1_intestine = intestinal abundance of UDP glucuronosyltransferase 1A1.

3.2.3. Synthesis of Constituent Analyses and Identification of Target Chemicals and Mixtures

The second objective of this study was to identify specific individual chemicals and mixtures in the Milwaukee Estuary that may require further effects-based characterization. As such, important cumulative effect/activity contributors identified through apical, non-apical, and pharmacological MCR evaluations were considered alongside results from predictive analyses to identify chemicals and mixtures that represent priority targets for further monitoring, risk assessment or ecotoxicological evaluation.

Evaluation of apical ecotoxicological potential (apical LoE; benchmark = fathead minnow 96-h LC50) identified one high priority mixture (PAHs; anthracene, benzo[a]pyrene, fluoranthene, phenanthrene, pyrene) and one medium priority mixture (sterols; β-sitosterol, cholesterol) within the Milwaukee Estuary. These findings are partially supported by previous prioritization efforts. For example, many of these PAHs have been identified as chemicals of potential concern in the Milwaukee Estuary ((companion study); (Baldwin et al. 2013)) and other Great Lakes watersheds (Corsi et al. 2019) based on detection frequency, environmental fate and/or ecotoxicological potential. Moreover, the two sterols have also been identified as potential high priority compounds in the Milwaukee Estuary, albeit with significant data limitations that could have influenced the assessment ((companion study)). However, a critical evaluation of these findings in the context of published ecotoxicological data demonstrates that some additional factors should be considered for mixture prioritization. For instance, PAHs and their mixtures have been shown to elicit embryotoxicity, cardiotoxicity, developmental abnormalities, decreased growth, and other apical effects in fish (Logan 2007; Le Bihanic et al. 2014) and invertebrates (Erickson et al. 1999; Ankley et al. 2003; Baldwin et al. 2017). Thus, these mixtures should be considered definitive high priority targets for effects-based monitoring and/or risk assessment within the Milwaukee Estuary. Conversely, sterols are naturally-derived, originating from decomposition of organic matter (Murtaugh and Bunch 1967; Mudge and Lintern 1999) and are not be expected to be inherently toxic. However, limited published studies have evaluated the ecotoxicities of these chemicals. Thus, sterol mixtures should be considered a medium priority target for further ecotoxicological evaluation, contingent on further evidence demonstrating significant risk potential in aquatic environments.

Evaluation of estimated bioactivity (non-apical LoE; benchmark = ToxCast ACC) identified one high priority mixture and five medium priority mixtures within the Milwaukee Estuary. The high priority mixture (ESR1_activation) was comprised of 17 diverse industrial chemicals, PAHs, PPCPs, pesticides, and flame retardants associated with bioactivities related to ERα activation. Medium priority mixtures were associated with bioactivities related to xenobiotic metabolism (CYP2B6_activation, NR1I2_activation), antioxidant response (NFE2L2_activation), lipid homeostasis (NR1I3_binding), and cellular stress responses (TP53_activation). Some of these findings are supported by whole mixture evaluations. For example, elevated endocrine-related bioactivity and/or in vivo effects were observed in all sites dominated by ESR1_activation mixtures, elevated PXR-related bioactivities and in vivo effects were observed in the site dominated by NR1I2_activation mixtures, and elevated PPAR bioactivity was observed at the site dominated by NR1I3_binding mixtures. Moreover, although these complex mixtures have not been experimentally evaluated, qualitative evidence demonstrates that activation of some of these target pathways could lead to adverse outcomes in various biota including hepatic steatosis (Angrish and Chorley 2021a; Angrish and Chorley 2021b; Angrish and Chorley 2021c), impaired population trajectories, altered reproductive behaviour or larval development, and/or impaired development of reproductive organs (Hutchinson and Villeneuve 2021). Thus, it is possible that these mixtures could elicit molecular effects that ultimately yield adverse outcomes in exposed aquatic organisms. As such, these six mixtures should be considered important targets for further evaluation, yielding more definitive characterizations of their cumulative sublethal or pathway-based effects in aquatic biota.

Evaluation of estimated pharmacological activity (pharmacological LoE; benchmark = Cmax) identified one high priority mixture (SLC_inhibition) within the Milwaukee Estuary. This mixture was comprised of six pharmaceuticals that inhibit solute carriers involved with neurotransmission, including sodium-dependent noradrenaline (SLC6A2), dopamine (SLC6A4) and serotonin (SL6A4) transporters. To date, this pharmaceutical mixture has not been formally evaluated in any published study. Thus, potential cumulative effects on aquatic organisms in the Milwaukee Estuary could not be directly determined. However, individually, some of these pharmaceuticals have been shown to elicit sublethal effects in aquatic biota. For example, venlafaxine has been shown to cause embryonic malformations, impair sensorimotor reflexes, alter gene and nuclear receptor expression, and trigger foot detachment in freshwater snails (Fong and Hoy 2012; Rodrigues et al. 2020). Similarly, exposure to citalopram has been shown to trigger foot detachment in freshwater snails (Fong and Hoy 2012) and impair feeding behaviour and induce oxidative stress in water fleas (Duan et al. 2022). Furthermore, some studies have shown that these pharmaceuticals can elicit behavioural effects in fish, impacting predation, stress response, and/or locomotion (Bisesi et al. 2014; Franco et al. 2019; Ziegler et al. 2020). However, evidence for these pharmaceutical-mediated behavioural effects varies amongst studies. Overall, the weight-of-evidence suggests that these solute carriers and their mixtures require further testing to discern potential sublethal and/or behavioural effects in aquatic biota. As such, they should be considered high priority targets for further ecotoxicological evaluation.

Random forest regression paired with weight-of-evidence evaluations identified three mixtures that were significant predictors of endocrine related effects and seven mixtures that were significant predictors of xenobiotic response related effects. Endocrine-related mixtures included sterols, nicotine and cotinine, antidepressants, and aromatic solvents, flavours and fragrances. Review of the published literature demonstrated that only a few of these compounds (e.g., galaxolide, nicotine, venlafaxine) have been shown to elicit endocrine-related effects in in vivo or in vitro assays (Kanungo et al. 2012; Overturf et al. 2015; Cavanagh et al. 2018). However, antidepressants, nicotine derivatives, and sterols have been shown to be suitable bioindicators for domestic wastewater exposure (Buerge et al. 2008; Schultz et al. 2010; Reichwaldt et al. 2017). Thus, it is possible that endocrine-related effects observed in study sites could be attributed to wastewater exposure (e.g., (Barber et al. 2015)), with the ‘endocrine-related mixtures’ identified here merely representing correlative bioindicators for more complex effluent exposure. Alternatively, the xenobiotic response-related mixtures tended to be complex, containing various industrial chemicals, PAHs, fuels, PPCPs, and flame retardants. Many of these mixtures were estimated to target various pathways linked to adverse outcomes (e.g., hepatic steatosis (Angrish and Chorley 2021a; Angrish and Chorley 2021b; Angrish and Chorley 2021c), altered reproductive behaviour or larval development (Hutchinson and Villeneuve 2021), and reproductive dysfunction and altered population dynamics (Santana Rodriguez et al. 2022; Villeneuve 2022)). Furthermore, individually, many of these mixture constituents (e.g., benzo[a]pyrene, fluoranthene, metolachlor, pyrene, etc.) have been shown to adversely impact reproduction, development, and morphology of exposed aquatic biota (Collier et al. 2013; Yang et al. 2021). Thus, there is some evidence that these xenobiotic response related mixtures may be of ecotoxicological concern. Overall, as this set of analyses prioritized mixtures using predictive relationships, causal links could not be established between mixture exposure and measured effects. However, the weight-of-evidence indicates that there may be some relationship between occurrence of these mixtures and sublethal endocrine and/or xenobiotic related responses. Therefore, these nine mixtures should be considered targets for further ecotoxicological testing and monitoring efforts, better characterizing how mixture occurrence may relate to in vivo and in vitro effects observed in the Milwaukee Estuary.

Along with mixtures, MCR and predictive analyses identified individual chemicals that significantly contributed to cumulative effects at the study sites. Overall, 16 individual compounds were highlighted as significant mixture effect contributors under apical, non-apical, pharmacological, and predictive LoE. These chemicals included three industrial compounds (4-tert-octylphenol, anthraquinone, bisphenol A), two WWIs (β-sitosterol, cholesterol), five PPCPs (acetaminophen, desvenlafaxine, metformin, methotrexate, nicotine, tramadol), two PAHs (benzo[a]pyrene, fluoranthene), one pesticide (metolachlor), and two fire retardants (tributyl phosphate, tris(2-butoxyethyl)phosphate). Interestingly, many of these compounds were also highlighted as high or medium priority chemicals in a companion study focused on prioritization of individual contaminants in the Milwaukee Estuary based on detection characteristics, environmental fate properties, ecotoxicity, and predictive relationships with measured effects (companion paper). For example, fluoranthene, benzo[a]pyrene, cholesterol, and β-sitosterol were all flagged as high priority chemicals, whereas anthraquinone, acetaminophen, desvenlafaxine, metformin, metolachlor, nicotine, and tramadol were considered medium priority chemicals. Others have also identified anthraquinone, bisphenol A, benzo[a]pyrene, fluoranthene, metolachlor, tributyl phosphate, and tris(2-butoxyethyl)phosphate as chemicals of potential ecotoxicological concern across the Great Lakes (Hull et al. 2015; Baldwin et al. 2016; Corsi et al. 2019) and other freshwater systems (Kostich et al. 2017; Deere et al. 2021). As such, the weight-of-evidence indicates that these 12 chemicals should be considered higher priority compounds for further evaluation. Conversely, the remaining two chemicals (4-tert-octylphenol, methotrexate) have rarely been identified as priority chemicals in single-chemical prioritization efforts. Thus, they may represent important targets for further studies better characterizing how they may contribute to mixture effects in the Milwaukee Estuary.

Overall, this constituent-based mixture prioritization analysis represents an important complement to other types of prioritization approaches (e.g., (Corsi et al. 2019), (companion paper)), in that it identifies chemicals that may be overlooked when they are evaluated based only on individual ecotoxicological or physicochemical properties. For example, if multiple chemicals have similar MOA or target similar biological pathways there is the potential for ecotoxicological effects to occur at concentrations that would not trigger high priority classifications on an individual basis. Thus, output from this mixture-based prioritization could be applied to supplement single-compound prioritization results, either providing additional targets for ecotoxicological evaluation or increasing the weight-of-evidence supporting the identification of high- and low-priority compounds in watersheds of interest.

4. Conclusions

This study employed two different traditional approaches to mixture analysis (holistic and constituent-based) to evaluate complex mixture effects in the Milwaukee Estuary. Integrated effects-driven methods were used to assess whole mixtures based on chemical composition and measured responses in in vitro and in vivo assays. Computational tools and techniques were used alongside predictive analyses to identify individual constituents and simplified mixtures representing significant contributors to mixture activity defined under diverse effect categories (apical, non-apical, pharmacological). Whole mixture evaluations were used to characterize Milwaukee Estuary sites, identifying those that could be targeted for complex urban/industrial (MEC, MET, UCJ, KKL) and PPCP (JIP, MIM, MIP, MIE) mixture evaluation, low priority sites (CCM, MIN), and a relatively unimpacted site (MEF) that could be used as a field reference in future monitoring studies. Constituent-based analyses narrowed down the 144 detected mixtures and 77 detected chemicals into 14 mixtures and 16 individual compounds representing priority targets for ecotoxicological testing, effects-based monitoring, or risk assessment. Overall, this study comprehensively assessed potential mixture toxicity in the Milwaukee Estuary, demonstrating how NAMs and classical mixture assessment tools could be integrated to expansively characterize complex mixture toxicity in a target watershed.

Table 4.

Chemicals identified as important contributors to mixture effects in the Milwaukee Estuary (2017 – 2018). Important drivers of mixture effect were identified using apical, non-apical, and pharmacological effect data and cumulative ratio (CR) analyses. Individual chemicals representing important predictors of endocrine-related (ER) and xenobiotic response-related (XR) effects in in situ caged fish studies were identified using random forest regression. High, medium, and low priority chemicals were identified based on the consensus of predictions generated from component-based and random forest analyses.

CAS	Chemical	CR Analyses			Effect Prediction		Chemical Prioritization
		LoE(s)	CR < 5 (n site-years)	TQ ≥ 10−2 (n site-years)	LoE	Proportion of Selected Effects	Priority Group	Supporting LoEs
50-42-8	Benzo[a]pyrene	Non-Apical	6	1	-	<0.2	High Priority	Non-Apical
80-05-7	Bisphenol A	Non-Apical	3	2	-	<0.2	High Priority	Non-Apical
206-44-0	Fluoranthene	Apical	17	3	-	<0.2	High Priority	Apical
93413-62-8	Desvenlafaxine	Pharmacological	8	8	ER	0.2	High Priority	Pharmacological; ER Effect Prediction
51218-45-2	Metolachlor	Non-Apical	8	3	-	<0.2	High Priority	Non-Apical
54-11-5	Nicotine	Pharmacological	7	6	-	<0.2	High Priority	Pharmacological
657-24-9	Metformin	Pharmacological	2	0	XR	0.25	Medium Priority	Pharmacological; XR Effect Prediction
140-66-9	4-tert-octylphenol	Non-Apical	1	1	-	-	Medium Priority	Non-Apical
83-46-5	β-sitosterol	Apical	4	0	ER XR	0.4 0.25	Medium Priority	Apical; ER Effect Prediction; XR Effect Prediction
57-88-5	Cholesterol	Apical	2	0	ER	0.2	Medium Priority	Apical
59-05-2	Methotrexate	Non-Apical	1	1	-	<0.2	Medium Priority	Non-Apical
126-73-8	Tributyl phosphate	Non-Apical	1	1	-	<0.2	Medium Priority	Non-Apical
78-51-3	Tris(2-butoxyethyl)phosphate	Non-Apical	1	1	-	<0.2	Medium Priority	Non-Apical
84-65-1	Anthraquinone	Effect Prediction	-	-	XR	0.44	Medium Priority	XR Effect Prediction
27203-92-5	Tramadol	Effect-Prediction	-	-	XR	0.25	Medium Priority	XR Effect Prediction
103-90-2	Acetaminophen	Effect Prediction	-	-	XR	0.25	Medium Priority	XR Effect Prediction

Data Availability Statement:

Data generated and analyzed in this study are available in this published article and its supplementary files. Data can also be accessed at https://cfpub.epa.gov/si/.