The Alginate Immobilization of Metabolic Enzymes Platform Retrofits an Estrogen Receptor Transactivation Assay With Metabolic Competence

Abstract

The U.S. EPA Endocrine Disruptor Screening Program utilizes data across the ToxCast/Tox21 high-throughput screening (HTS) programs to evaluate the biological effects of potential endocrine active substances. A potential limitation to the use of in vitro assay data in regulatory decision-making is the lack of coverage for xenobiotic metabolic processes. Both hepatic- and peripheral-tissue metabolism can yield metabolites that exhibit greater activity than the parent compound (bioactivation) or are inactive (bioinactivation) for a given biological target. Interpretation of biological effect data for both putative endocrine active substances, as well as other chemicals, screened in HTS assays may benefit from the addition of xenobiotic metabolic capabilities to decrease the uncertainty in predicting potential hazards to human health. The objective of this study was to develop an approach to retrofit existing HTS assays with hepatic metabolism. The Alginate Immobilization of Metabolic Enzymes (AIME) platform encapsulates hepatic S9 fractions in alginate microspheres attached to 96-well peg lids. Functional characterization across a panel of reference substrates for phase I cytochrome P450 enzymes revealed substrate depletion with expected metabolite accumulation. Performance of the AIME method in the VM7Luc estrogen receptor transactivation assay was evaluated across 15 reference chemicals and 48 test chemicals that yield metabolites previously identified as estrogen receptor active or inactive. The results demonstrate the utility of applying the AIME method for identification of false-positive and false-negative target assay effects, reprioritization of hazard based on metabolism-dependent bioactivity, and enhanced in vivo concordance with the rodent uterotrophic bioassay. Integration of the AIME metabolism method may prove useful for future biochemical and cell-based HTS applications.

Substantial progress has been made in high-throughput toxicity testing in the U.S. Environmental Protection Agency (EPA) ToxCast (Kavlock et al., 2012), and inter-agency Tox21 (Tice et al., 2013), programs to evaluate the impact of chemicals on human health and the environment. The Endocrine Disruptor Screening Program (EDSP) (EPA US 2014, 2017) utilizes ToxCast/Tox21 data to evaluate the biological effects of potential endocrine active substances (EAS) on estrogen (Browne et al., 2015; Judson et al., 2015), androgen (Kleinstreuer et al., 2017, 2018), thyroid (Deisenroth et al., 2020; Friedman et al., 2016; Olker et al., 2019; Paul-Friedman et al., 2019; Wang et al., 2018), and steroidogenesis pathways (Haggard et al., 2018, 2019; Karmaus et al., 2016). A potential limitation to the use of in vitro assay data in regulatory decision-making is the lack of coverage for xenobiotic metabolic processes occurring within an organism. Indeed, hepatic- and peripheral-tissue metabolism can yield metabolites that exhibit greater activity than the parent compound (bioactivation) or are inactive (bioinactivation) for a given biological target. Interpretation of biological effect data for both putative EAS, as well as other chemicals, screened in high-throughput screening (HTS) assays may benefit from incorporation of xenobiotic metabolic competence to decrease uncertainty associated with predicting potential human health hazards from in vitro assay data (Jacobs et al., 2008, 2013; OECD, 2014).

Methods to incorporate xenobiotic metabolism into in vitro assays can generally be classified as intracellular or extracellular. Intracellular approaches are cell-based and encompass basal, induced, or engineered metabolism. Primary hepatocytes or hepatic cell lines (Gripon et al., 2002; Soldatow et al., 2013), recombinant cell lines engineered to express cytochrome P450 enzymes (Gomez-Lechon et al., 2017), or transfection of modified mRNA for transient cytochrome P450 expression (DeGroot et al., 2018) have all been utilized to capture features of intrinsic liver metabolism. Extracellular approaches, defined by metabolism external to the cellular environment, have utilized recombinant xenobiotic-metabolizing enzymes (Yu et al., 2018), or hepatic subcellular fractions such as microsomes and postmitochondrial supernatant (S9) (Mollergues et al., 2017; van Vugt-Lussenburg et al., 2018). Hepatic S9 contains both phase I and II metabolic enzymes present in the microsomal (eg, cytochrome P450s, uridine 5′-diphospho-glucuronosyltransferases, carboxylesterases) and cytosolic (eg, sulfotransferases, glutathione S-transferases, methyltransferases, N-acetyl transferases, xanthine oxidase, aldehyde oxidase) fractions (Parmentier et al., 2006). Historically, rodent and human S9 metabolizing systems are used in the Ames test and are an accepted method for non-quantitative mutagenicity screening (Callander et al., 1995; Hakura et al., 2003; Maron and Ames, 1983; OECD, 2020). As a result, direct application of exogenous S9 has been used to examine the impact of metabolism on estrogen- (Charles et al., 2000; Mollergues et al., 2017; van Vugt-Lussenburg et al., 2018), androgen- (de Rijke et al., 2013; Kuuranne et al., 2008; Mollergues et al., 2017; van Vugt-Lussenburg et al., 2018), and thyroid-active substances (Taxvig et al., 2011). However, despite the success of S9 addition as an accessory approach to incorporating metabolism into some in vitro endocrine assays, microsomal and S9 preparations have been observed to contribute to cytotoxicity (Bimboes and Greim, 1976; Cox et al., 2016), enhance chemical protein binding that may alter toxicological activity (Heringa et al., 2004; Hoogenboom et al., 2002), or cause technical interference in certain image- or cell-based assays (Dreier et al., 2002).

One strategy to mitigate S9-associated assay interference is physical separation of S9 from the target assay components using hydrogel encapsulation. Alginate is a polysaccharide extracted and processed from brown seaweed that forms a gel-like structure in the presence of divalent cations (Thu et al., 1996). Due to its favorable biocompatibility, alginate has been used in a variety of pharmaceutical and biomedical applications to encapsulate cells (Lee and Mooney, 2012; Smidsrod and Skjak-Braek, 1990). For example, alginate encapsulation of human hepatocytes has been used to evaluate intrinsic clearance rates in the liver for computational modeling (Phillips et al., 2018), and for identifying promutagenic effects of potential carcinogenic substances (Vian et al., 2002). An HTS-compatible system has been developed using live cells encapsulated in alginate on plastic pillar inserts for 96-well microplates (Lee et al., 2008), and scaled to 384-well microplates to evaluate the impact of recombinant liver enzymes on model compounds known to affect cell viability (Yu et al., 2018). Alternatively, an encapsulation approach that replaces cells with S9 has been used to evaluate cytotoxicity of reactive metabolites while avoiding S9 generated toxicants such as lipid peroxides (Yamamoto et al., 2011), suggesting hepatic S9 encapsulation in alginate hydrogel is a viable approach for incorporating metabolism into HTS assays.

The objective of this study was to integrate features of alginate encapsulated hepatic S9 into an HTS-compatible system to screen for xenobiotic metabolism-based effects on EAS. The Alginate Immobilization of Metabolic Enzymes (AIME) platform consists of custom 96-well microplate lids containing solid supports attached to encapsulated hepatic S9-alginate microspheres. The optimized platform uses rat hepatic S9 to evaluate phase I xenobiotic metabolism while minimizing cytotoxic and assay interference effects. Deployment to an estrogen receptor transactivation (ERTA) assay with a test set of compounds previously identified by estrogen receptor quantitative structure-activity relationship (ER-QSAR) analysis to generate ER-active metabolites (Pinto et al., 2016) demonstrates utility for identification of false-positive and false-negative target assay effects, reprioritization of hazard based on metabolism-dependent bioactivity, and enhanced in vivo concordance with the rodent uterotrophic bioassay. The results support application of the AIME method for enhancing in vitro predictions of estrogen-active substances and may be useful for future biochemical and cell-based HTS applications.

MATERIALS AND METHODS

Materials.

Ultra-pure, medium viscosity sodium alginate (UP MVG) was purchased from NovaMatrix (Sandvika, Norway). Barium chloride (99.9% purity), poly-L-lysine (PLL) (molecular weight: 70 000–150 000), NADP+, glucose-6-phosphate, magnesium chloride, and DMSO (molecular biology grade) were from MilliporeSigma (St. Louis, Missouri). Phenobarbital/β-naphthoflavone-induced male Sprague Dawley rat hepatic S9 was purchased from Molecular Toxicology (Boone, North Carolina).

Design and Fabrication of Custom AIME Lids and Trough Plates.

Design, engineering, and production of CAD files for the AIME 96-well lids and trough plates were performed by Innovative Design Engineering (Raleigh, North Carolina). Lids and trough plates were produced by custom injection molding (3D Systems, Rock Hill, North Carolina) using a clear, general-purpose polystyrene resin (Styron 666H). The final AIME lids were cleaned and the surface functionalized for enhanced microsphere adhesion through gas plasma activation (PVA Tepla America, Inc., Corona, California). Trough plates were individually packaged and sterilized by ethylene oxide processing prior to use.

AIME Platform Setup.

The AIME procedure is a modification of protocols previously reported for encapsulating cells in an alginate matrix (Lee et al., 2008, 2013), and consists of a preparative “pre-stamping” step using a barium chloride/PLL solution followed by “stamping” the lid in alginate-hepatic S9 to form the metabolically active microspheres. In the prestamping step, AIME lids were placed into a sterile, 96-well plate containing 150 μl per well of a sterile 1:2 (vol/vol) solution of 100mM barium chloride and 0.01% PLL. The lids were incubated in solution for 10 min at room temperature, then allowed to air dry for a minimum of 10 min.

To prepare the stamping solution of alginate-hepatic S9, a 2.5% solution of sodium alginate was first prepared in sterile, deionized water and allowed to dissolve for 24–48h at 4°C, with shaking, as required for complete dissolution. Phenobarbital/β-naphthoflavone-induced male Sprague Dawley rat hepatic S9 was thawed in a 37°C water bath for 4 min and diluted to 20 mg/ml in chilled 0.15M potassium chloride. The diluted hepatic S9 was mixed with the 2.5% alginate solution to achieve a final hepatic S9 concentration of 10% (vol/vol). Two milliliters of the alginate-hepatic S9 solution was loaded per trough into a chilled (0°C) 12-trough plate and centrifuged at approximately 700 rpm for 1 min to evenly distribute the alginate-hepatic S9 solution within each trough. Empty alginate control lids were prepared as described for the metabolically active lids by substituting sterile water for the hepatic S9. Prestamped AIME lids were sequentially incubated for 1 min in the chilled alginate-hepatic S9 solution, moved to new trough plates (1 trough plate per lid) containing 100mM barium chloride for 1 min, then moved to a final trough plate containing a 0.01% PLL rinse for 1 min. Once stamping was completed, the AIME lids were used immediately for biochemical or cell-based assays. Preparation of the AIME lids for use in cell-based assay was performed using aseptic technique.

Evaluation of Cytochrome P450 Metabolism.

Metabolic activity was evaluated across a reference set of probe substrates recommended for in vitro study of human cytochrome P450s (Bjornsson et al., 2003) to determine performance of rat S9 in the alginate microspheres. Phenacetin (CASRN 62–44-2), diclofenac sodium salt (CASRN 15307–79-6), dextromethorphan hydrobromide monohydrate (CASRN 6700–34-1), chlorzoxazone (CASRN 95–25-0), bupropion hydrochloride (CASRN 31677–93-7), and carbofuran (CASRN 1563–66-2) were obtained from MilliporeSigma (St. Louis, Missouri) (Supplementary Table 1). Chemical stocks were prepared in DMSO and stored at −80°C. Estrogen-stripped DMEM (without phenol red, low glucose) supplemented with 4mM L-glutamine, 10mM HEPES, and 1% charcoal/dextrantreated fetal bovine serum (FBS) was charged with NADPH using an NADPH regeneration system (NRS) containing 1.3mM NADP+, 3.3mM glucose-6-phosphate, 3.3mM magnesium chloride, and 0.3 U/ml glucose-6-phosphate dehydrogenase (Worthington Biochemical Corporation, Lakewood, New Jersey). Following NADPH generation, the charged medium was diluted 1:11.4 using additional estrogen-stripped medium without an NRS. Chemical stocks were diluted into the estrogen-stripped medium to a final DMSO concentration of 0.2% (vol/vol) and added to empty 96-well microplates (120 μl/well). Final test concentrations were as follows: phenacetin (80 μM), diclofenac sodium salt (100 μM), dextromethorphan hydrobromide monohydrate (80 μM), chlorzoxazone (500 μM), bupropion hydrochloride (500 μM), and carbofuran (100 μM). Freshly prepared AIME lids were added to the assay plates (except for the 0-h timepoint microplate), and the plates incubated at 37°C and 5% CO₂ for 2, 4, and 8h. At the termination of each timepoint, AIME lids were removed and 100 μl from each well stored at −80°C until analytical analysis.

Liquid Chromatography-Tandem Mass Spectrometry.

Substrate standards, metabolite standards, and internal standards are provided (Supplementary Table 2). Calibration standards and quality control (QC) samples were matrix-matched utilizing the diluted, charged estrogen-stripped medium to match initial sample volume. Standards, controls, and samples were prepared by spiking internal standards into 1% formic acid in acetonitrile or methanol at 3× the sample volume, vortexed to mix, and centrifuged at 10 000 rpm. Supernatant was filtered through Phree Phospholipid Removal, 30 mg/well, 96-well plates (Phenomenex, Torrance, California) using a negative-pressure vacuum apparatus at 3–5 psi (0.2–0.3 bar) and transferred to liquid chromotography (LC) vials for analysis. Samples were prepared in batches with solvent blank, method blank, 8 matrix-matched calibration standards, and matrix-matched QC samples at low and high concentrations.

Instrumental analysis was performed on an Acquity Ultra Performance LC paired with a Quattro Premier XE triple quadrupole mass spectrometer (MS) (Waters Corporation, Milford, Massachusetts). LC separation was achieved using a Kinetex C18 LC column (100×2.1mm, 2.6 μm, 100 å; Phenomenex, Torrance, California) under gradient elution using either 95/5 water and methanol with 2.5mM ammonium acetate (A1) and 95/5 methanol and water with 2.5mM ammonium acetate (B1), or 0.1% formic acid in water (A2) and 0.1% formic acid in acetonitrile (B2) depending on the method. LC and MS parameters are listed (Supplementary Table 3).

Integration, calibration, and quantitation of samples were performed using Quanlynx software (Waters Corporation, Milford, Massachusetts). Calibration data were fit to quadratic curves using 1/x² weighting. Batch results were accepted based on the following criteria: (1) calibration curves exhibited a correlation coefficient of >0.98 (0.99 for accuracy/precision data), (2) standards and QC accuracy tolerance ≤ 20% (30% at lower limit of quantification [LLOQ]), (3) QC precision expressed as percentage of relative standard deviation (% RSD) ≤15% (20% LLOQ), and (4) solvent and method blank response were free of target analyte. Performance measurements were calculated for accuracy and precision of target analytes (Supplementary Table 4). All samples where no analyte was detected were set to zero for statistical purposes.

ERTA Assay Chemical Set.

Reference chemicals for the VM7Luc ERTA assay were identified from OECD Test Guideline 455 (OECD, 2016) and 457 (OECD, 2012) (referenced hereafter as “test guideline”): one assay reference standard (17β-estradiol [E2]), 6 ER agonists (17α-ethinylestradiol, bisphenol A, coumestrol, ethylparaben, genistein, daidzein); 3 ER antagonists (tamoxifen, 4-hydroxytamoxifen, fulvestrant); 3 agonist negative controls (atrazine, corticosterone, spironolactone); and 1 antagonist negative control (resveratrol). The organochlorine pesticide methoxychlor (MXC) was included as a prototypical chemical requiring metabolic activation to produce the potent estrogenic metabolite 2,2-bis(4-hydroxyphenyl)-1,1,1-trichloroethane (HPTE) (Bulger et al., 1978a,b,c; Gaido et al., 1999; Ousterhout et al., 1981; Shelby et al., 1996). Both MXC and HPTE were included to monitor the metabolism-dependent range of MXC ERTA bioactivity. Thirty-four parent chemicals predicted to have more potent estrogenic metabolites (metabolism positive test set), and 14 parent chemicals predicted to have minimal estrogenic activity and no evidence of conversion to estrogenic metabolites (metabolism negative test set) were selected from prior literature (Pinto et al., 2016). The complete list of 63 chemicals with DSSTox substance identifier (DTXSID), CASRN, chemical name, vendor, lot, and purity are found in Supplementary Table 1. Test chemicals were solubilized in DMSO to a final concentration of 100mM, or up to the limit of solubility, and stored at −80°C. Plate-based controls run on each assay plate were as follows: E2 (100 pM) for ERTA, MXC (10 μM) for metabolismdependent ERTA, and DMSO (0.2%) as solvent control. The number of technical replicates per plate for control and test chemicals is indicated on the plate map (Figure 3B).

Deployment of the Aime Assay to An Erta Assay.

The VM7Luc4E2 (formerly BG1Luc4E2) (Li et al., 2014; Rogers and Denison, 2000) cell line was kindly provided by Dr Michael S. Denison (University of California – Davis). Cells were maintained with routine passaging in Minimum Essential Medium alpha (MEM α) supplemented with 10% FBS (Atlanta Biologicals, Flowery Branch, Georgia) and 312 μg/ml G418 sulfate (Thermo Fisher Scientific, Waltham, Massachusetts) in an incubator set at 37°C and 5% CO₂. For detection of ER agonists, cells were stripped of exogenous estrogens via a 5-day protocol. Initially, 3.0 × 10⁶ cells were seeded into T-175 flasks (Corning Inc., Corning, New York) in MEM α without G418 sulfate and allowed to attach for 24 h. The cells were rinsed twice with phosphate-buffered saline, followed by the addition of estrogen-stripped medium (DMEM (phenol red-free, low glucose)) supplemented with 4mM L-glutamine, and 5% charcoal/dextran-treated FBS. The estrogen-stripped medium was changed daily for 3 days. On Day 4, estrogen-stripped VM7Luc4E2 cells were seeded into white 96-well microplates at 7.5 × 10⁴ cells per well using a MultiFlo FX automated reagent dispenser (BioTek, Winooski, Vermont) and acclimated for 18–24 h in an incubator set at 37°C and 5% CO₂. Metabolism of test compounds was performed in assay medium containing estrogen-stripped DMEM (phenol red-free, low glucose) supplemented with 4mM L-glutamine, 10mM HEPES, and 1% FBS and charged with an NRS for 10 min at room temperature to allow for formation of NADPH. Prior to use in the assay, the charged medium was diluted 1:11.4 in additional uncharged assay medium.

Chemical source plates were prepared in Echo qualified 384-well polypropylene microplates (Labcyte, San Jose, California) using an Echo 550 acoustic liquid handler (Labcyte, San Jose, California) and Certus Flex automated liquid dispensing system for DMSO backfill (Fritz Gyger AG, Switzerland). Source plates were sealed and stored at room temperature (20°C–25°C), protected from light, for the duration of the screening experiments. For the agonist assay, test chemicals were dispensed into white 96-well microplates (Greiner Bio-One, Monroe, North Carolina) using an Echo 550 liquid handler (Labcyte, San Jose, California). A dilution series (0.002–200 μM) was generated by dispensing variable nanoliter volumes of test chemical and backfilling with a complementary volume of DMSO using Echo Cherry Pick software (v.1.6.2) to a final total dispensed volume of 250 nl. A volume of 125 μl uncharged (column 1) or charged (columns 2–12) assay medium was added to each plate (0.2% DMSO concentration) using a MultiFlo FX automated reagent dispenser (BioTek, Winooski, Vermont). A freshly prepared AIME lid containing no microspheres in column 1 and metabolically inactive (water/alginate) or active (S9/alginate) microspheres in columns 2–12 was added to each assay plate for 2h at 37°C and 5% CO₂. The AIME lids were removed and 100 μl conditioned medium, containing parent compounds and metabolites, transferred to aspirated assay plates containing the estrogen-stripped VM7Luc4E2 cells using a ViaFlo 96-channel semi-automated pipette (Integra Bioscience, Hudson, New Hampshire). Following a 24-h incubation period at 37°C and 5% CO₂, the conditioned medium was aspirated from the assay plates and replaced with fresh assay medium (120 μl) containing a 1:6 dilution of CellTiter-Fluor cell viability reagent (GF-AFC substrate plus assay buffer) (Promega, Madison, Wisconsin) to measure treatment-dependent cytotoxicity. The plates were incubated at 37°C and 5% CO₂ for 30 min, followed by a fluorescent endpoint scan in a multi-mode CLARIOstar microplate reader (BMG Labtech Inc, Cary, North Carolina). For ER transactivation, 60 μl/well of reconstituted Bright-Glo reagent (Promega, Madison, Wisconsin) was subsequently added, and plates incubated for 3 min at room temperature prior to reading the luminescent endpoint on the CLARIOstar microplate reader. A total of 3 experimental replicates were performed for each chemical (n=3).

Toxcast Pipeline.

Raw concentration-response data with assay endpoint IDs (AEID) 2486 (ERTA luciferase; metabolism negative), 2484 (ERTA luciferase; metabolism positive), 2487 (cytotoxicity; metabolism negative), and 2485 (cytotoxicity; metabolism positive) (additional assay information in Supplementary Table 17) were fit and analyzed using the ToxCast Data Analysis Pipeline (TCPL v.2.0.1) (Filer et al., 2017). Well-level data (rval) were obtained as either raw luminescence units for the ERTA luciferase assay, or raw fluorescence intensity units for the cytotoxicity assay, and normalized to plate-level controls. For the luciferase assay, rvals were normalized as percent activity of positive control using the following equation,

$$resp=100\frac{(rval-bval)}{pval}$$ (1)

where resp is the normalized response variable for fitting, bval is the plate-level mean of the NRS supplemented DMSO control, and pval is the plate-level median background-corrected 17β-estradiol positive control (POS) under standard VM7Luc ERTA assay conditions (OECD, 2012) (no alginate microsphere; no NRS) where pval = median(rval_POS,VM7 – mean(rval_DMSO,VM7)). Comparatively, the raw fluorescence units from the cytotoxicity assay were normalized as a log₂ fold change using the following equation,

$$resp=-lo{g}_2\left(\frac{rval}{bval}\right)$$ (2)

where resp is the response variable to be fit and bval is the mean of the NRS supplemented DMSO control. Finally, the negative of equation 2 allowed for TCPL-required fitting in the positive direction.

Each chemical concentration-response was fit to 3 models, constant (zero parameters), hill (3 parameters), and gain-loss (5 parameters), with the winning model determined through the minimum of the Akaike Information Criteria (AIC). Finally, for a given assay endpoint, a chemical hit was assigned if the winning model structure was either a hill or gain-loss function with a maximum median response value and fitted top of the curve (tp) exceeding a threshold of 10 times the baseline median absolute deviation (bmad). Here, bmad represents the median absolute deviation of the normalized NRS supplemented DMSO control resp values across every plate of the assay.

Uncertainty Quantification Analysis.

To assess TCPL model fitting uncertainty, smooth nonparametric bootstrapping analysis of the AIME data using ToxBoot (v0.1.2) was conducted (Watt and Judson, 2018). Resampling of each response value (resp) was generated by adding zero-centered noise based on a normal distribution, N(0, bmad), and generating 1000 bootstrapped samples. TCPL nonlinear regression was performed for every bootstrapped sample with winning model (lowest AIC), and subsequent model parameters calculated and stored. In addition to the fitted parameters from each nonlinear regression such as AC₅₀ and maximum efficacy, the area under the fitted curve (AUC) was also determined for every bootstrapped sample. For a resample with a hit call of 1 (yes), the AUC for the winning model parameters was determined using the trapezoidal rule. A resampled model with a hit call of 0 (no) was assigned an AUC of 0. The resulting distributions for AC₅₀, max efficacy, and AUC allowed for characterization of model uncertainty for each chemical-AIME assay mode pair. The hit probability (%hitc) was calculated as the percentage of hit calls (hitc) made across the total number of bootstrapped samples for AUC (Supplementary Tables 7–14). All data were plotted, visualized, and analyzed using GraphPad Prism v8.1.2 (GraphPad Software, San Diego, California) and RStudio (Integrated Development for R. RStudio, Inc., Boston, Massachusetts).

Data Analysis.

Assay summary performance metrics for the control compounds were determined for inter-assay variability (coefficient of variation [CV]) and screening quality (Z′-factor) (Zhang et al., 1999) across metabolism modes and cell culture medium conditions. CV was determined at the plate-level for all 3 control compounds (DMSO, E2, and MXC) according to pair-wise assay conditions (±) metabolism and (±) NRS. Likewise, the Z′-factor was calculated from negative (DMSO) and positive (E2) controls under the same conditions. The AIME-ERTA assay was not run in metabolism positive mode without NRS, so this pair was excluded from analysis (Table 1).

Designation of hit calls for metabolism-dependent effects in the ERTA assay was evaluated using 2 approaches: (1) the binary TCPL yes/no (1/0) hit call determinations from the raw data curve fits and (2) confidence intervals calculated from the difference in bootstrapped AUC sampling data (Table 2). Binary hit calls were established from TCPL fits in metabolism negative (AEID 2486) and positive (AEID 2484) modes. Directional shifts for metabolism-dependent effects were determined by the difference between positive versus negative modes (ΔHitc_ER) where bioinactivation = −1, no change = 0, and bioactivation = 1. Classification of biotransformation using this approach required parent compound effects to be different than the metabolites as a function of the TCPL cutoff (10*bmad). This approach was sufficient to capture a subset of false-negative (bioactivated) and false-positive (bioinactivated) chemicals that had marked changes around the cutoff, but was inherently susceptible to variation with changes in the semi-arbitrary cutoff and insufficient to classify all metabolism-dependent effects.

Alternatively, the difference in the mean AUC between metabolism positive versus negative modes (ΔAUC) for the bootstrapped sampling data defined the directional shift trends where bioinactivation < 0, no change = 0, and bioactivation > 0. Confidence intervals (CI) for bootstrap estimates of standard error were used to distinguish the overlap in sampling variability between the 2 assay modes and were calculated with the following equation,

$$\text{CI}=\left({\mu }_p-{\mu }_n\right)\pm q\times \sqrt{{\sigma }_p^2+{\sigma }_n^2}$$ (1)

where CI is the confidence interval, μ_p and μ_n are the mean ERTA AUC signal in metabolism positive and negative modes, q is the quantile of the standard normal distribution, and σ_p and σ_n are the standard deviation for the ERTA AUC signal in metabolism positive and negative modes. A quantile value of 3 was applied for a desired confidence interval of 99.9%. A CI range < 0 indicated a separation of signal bands corresponding to significant bioinactivation, whereas a CI range of > 0 corresponded to significant bioactivation. Instances where the CI range overlapped with 0 indicated no significant metabolic effect. This approach identified additional compounds with high confidence that exhibited marked changes in metabolism-dependent bioactivity, but otherwise displayed no directional shift via conventional TCPL hit calls. Although not strictly pertinent to the identification of false-positive and false-negative hits in the ERTA assay, the relative bioactivity changes were important for chemical prioritization. In combination, TCPL hit calls and ΔAUC with CI enabled quantitative classification of metabolism-dependent biotransformation (activation and inactivation) in the AIME-ERTA assay.

For comparative in vitro to in vivo analysis (Table 2), ToxCast ER Model scores (Browne et al., 2015; Judson et al., 2015) and the number of guideline-like uterotrophic studies (GL-UT) (Kleinstreuer et al., 2016) were examined for all 63 chemicals in the study. A total of 222 GL-UT studies were identified for 24 of the chemicals (Supplementary Table 15). Binary yes/no (1/0) hit calls and weight-of-evidence (GL_WoE) were determined. AIME – VM7Luc ERTA concordance to the uterotrophic data was compared across metabolism negative (Met_Neg) and positive (Met_Pos) assay modes using TCPL hit calls. The change in concordance (ΔMet) is noted as increased (1), no change (0), or decreased (−1) predictivity. Data not available (NA) or equivocal (EQUIV) findings are documented.

RESULTS

Design of the AIME Platform

The AIME platform consists of a custom polystyrene lid with 96 solid support pegs designed to be compatible with all ANSI/SLAS standard microplates (Figure 1A). Each peg of the AIME lid can contain a single microsphere of alginate matrix mixed with hepatic S9 (Figure 1B) and reach a standard depth that provides sufficient distance to the well bottom for cell-free or cell culture applications (Figure 1C). Preparation of the AIME lids is similar to the process previously described (Lee et al., 2008) with important distinctions. First, the AIME lids were gas plasma treated to functionalize the polystyrene surface and enhance the attachment of the alginate-hepatic S9 microsphere. This eliminated the need for an initial Matrigel stamping step. Second, following stamping with the alginate-hepatic S9, the AIME lids were sequentially rinsed with barium chloride and PLL solutions. This procedure followed from protocols used to produce individual alginate microspheres (Leick et al., 2011; Yamamoto et al., 2011). The barium chloride rinse provided divalent cations necessary for alginate gelling and also produced a mechanically stronger cross-link than calcium for S9 encapsulation on the AIME lids (Mørch et al., 2006). The PLL rinse was added to control the porosity of the gelled alginate and assist in the retention of the xenobiotic-metabolizing enzymes within the alginate-S9 matrix (Goosen et al., 1985; Vandenbossche et al., 1993).

Figure 1.

Alginate Immobilization of Metabolic Enzymes (AIME) lid design. The AIME platform consists of a custom polystyrene lid with 96 solid supports functionalized for enhanced hydrogel attachment that is compatible with all ANSI/SLAS standard microplates (A). AIME lid with solid supports containing microspheres of hepatic S9 attached in an alginate matrix. The column on the left contains microspheres dyed with Ponceau-S to aid in visualization (B). Each solid support in a standard 96-well microplate provides sufficient distance between the end of the solid support and the well bottom for attachment of the microsphere without impacting the cell monolayer (C).

Characterization of the AIME Microsphere

Initial characterization of the AIME microsphere function focused on 2 key areas: (1) the ability of small molecules to freely enter the microsphere and undergo metabolism by the cytochrome P450 enzymes and (2) the duration of cytochrome P450 enzyme activity of the S9. To evaluate these features, a panel of 5 reference substrates (phenacetin, dextromethorphan, bupropion, diclofenac, and chlorzoxazone), known to be metabolized by human cytochrome P450 enzymes to specific and identifiable metabolites (Bjornsson et al., 2003), was incubated with the AIME lids for 0, 2, 4, and 8h and analyzed by LC-MS/MS. The results demonstrate that S9 functioned well within the alginate microsphere and formed detectable metabolites from the probe substrates (Figure 2; Supplementary Table 5). Fourteen percent of phenacetin (Figure 2A) was metabolized to acetaminophen by 2h, increasing to approximately 20% depletion at 4 and 8h. Dextromethorphan (Figure 2B) was extensively metabolized to dextrorphan, with 22%, 36%, and 40% depletion across the 2, 4, and 8-h timepoints. Bupropion (Figure 2C) was also highly metabolized with 30% and 36% depletion of the parent compound at 4 and 8h, respectively. Depletion of diclofenac (Figure 2D) at 2h was 22% with no further detectable decreases in the parent compound. However, formation of 4′-hydroxydiclofenac continued to increase from 0.97 μM at 2 h to 1.44 μM at 8h. Chlorzoxazone (Figure 2E) was metabolized to a minor extent over the 8-h time course, reaching maximum depletion of approximately 11%. The 6-hydroxychlorzoxazone metabolite was readily formed reaching a concentration of 6.4 μm by 2 h, and 13. 4 μM at 8 h. Carbofuran, a representative compound from the ToxCast chemical library, was tested at 100 μM to approximate a typical top concentration in a ToxCast assay (Figure 2F). At 2h, 16.6% of the compound had been depleted, increasing to 32.8% and 37.2% at 4 and 8h, respectively. Metabolite formation was relatively stable from 2 to 4h with a final formed concentration of 0.66 μM. The results demonstrate phase I cytochrome P450 oxidation was functional for encapsulated hepatic S9 on the AIME lids with test compounds successfully converted into predicted and quantifiable metabolites. Importantly, the production of metabolites rose rapidly during the first 2 h, then moderately continued throughout the duration of the 8-h time course, demonstrating sustained S9 activity throughout the encapsulation process and for several hours in cell culture medium.

Figure 2.

Evaluation of phase I cytochrome P450 metabolism in AIME microspheres. An array of cytochrome P450 substrates was evaluated for parent compound depletion and metabolite accumulation across a time series of 0, 2, 4, and 8h. Parent/metabolite pairs include phenacetin/acetominophen (A), dextromethorphan/dextrorphan (B), bupropion/hydroxybupropion (C), diclofenac/4′-hydroxydicofenac (D), and chlorzoxazone/6-hydroxychlorzoxazone (E). The final parent/metabolite pair, carbofuran/3-hydroxycarbofuran, was selected from the ToxCast library to demonstrate library compatibility (F). Data points are mean concentration (μM) plotted for each experimental replicates (n=3).

AIME Coupled to an ERTA Assay

Following confirmation of cytochrome P450 enzyme activity, the AIME platform was evaluated with the in vitro VM7Luc ERTA assay. The VM7Luc4E2 cell line utilizes a stably transfected ER responsive luciferase reporter to evaluate the transactivation potential of endogenous hERα, and to a lesser extent hERβ, in response to estrogenic test substances (OECD, 2016). Here, the method was multiplexed with a fluorescent viability endpoint to enable simultaneous detection of ER agonist potential (luminescence) and cytotoxicity (fluorescence) on a multi-mode microplate reader (Figure 3A). The assay was run in 96-well plate format with column 1 dedicated to VM7Luc ERTA baseline control conditions (OECD, 2012). E2 (100 pM) was run as the positive reference standard for the ERTA endpoint, DMSO (0.2%) as the solvent control, and MXC (10 μM) as the metabolism control (Figure 3B). MXC is a compound that undergoes sequential demethylation by cytochrome P450 enzymes to yield the estrogenic metabolite HPTE (Jacobs et al., 2008). All of the control wells in column 1 contained cell culture medium that conformed to the test guideline standard. For applying the AIME platform to the assay, column 2 evaluated the same set of controls, but did so with medium supplemented with the NRS. Initial experiments determined that more conventional NRS component concentrations were significantly cytotoxic, so the concentrations were optimized to mitigate cell-based cytotoxicity whilst maintaining cytochrome P450 metabolic activity (data not shown). Columns 3–12 of the assay plates were set up to test a 10-point titration series of test compound (0.002–200 μM) in the presence of cell culture medium supplemented with NRS (Figure 3B). An alternative dose spacing design was employed where concentration intervals were skewed toward the higher range to increase the number of data points that might reflect AIME-induced metabolite activity. The AIME assay was run in 2 modes simultaneously, specified by different lids: metabolism negative, where blank alginate-water microspheres were immobilized to lid pegs; and metabolism positive, where alginate-S9 microspheres were immobilized to lid pegs. The design enabled parallel detection of parent compound (metabolism negative) and metabolite(s) (metabolism positive) bioactivity within the same assay run of duplicate chemical-treated assay plates.

Figure 3.

Deployment of the AIME method to the VM7Luc estrogen receptor transactivation (ERTA) assay. The VM7Luc4E2 cell line utilizes a stably transfected ER responsive luciferase reporter to evaluate the transactivation potential of endogenous hERα, and to a lesser extent hERβ, in response to estrogenic chemicals. Multiplexed assay endpoints were ER transactivation (luminescence) and cytotoxicity (fluorescence) (A). The plate map for the assay was comprised of test guideline control conditions in column 1 (no NRS, no AIME microsphere), AIME control conditions in column 2 (NRS, AIME microsphere), and test compounds with alternative dose spacing in column 3–12 (0.002–200 μM). Control chemicals were ER positive control 17β-estradiol (E2; 100 pM), negative solvent control dimethyl sulfoxide (DMSO; 0.2%), and AIME metabolism control methoxychlor (MXC; 10 μM). Each experimental run consisted of duplicate treated plates run in metabolism negative and positive mode, respectively, for simultaneous evaluation of parent and metabolite effects (B). Performance of plate-based controls for column 1 E2 (C1_E2), column 2 E2 in metabolism negative mode (C2_E2(−)), and column 2 E2 in metabolism positive mode (C2_E2(+)). Plate-based normalization was calculated with column 1 E2 conditions (grey solid line), with mean responses for E2 metabolism negative (blue-dashed line) and E2 metabolism positive (red-dashed line) modes shown (C). Performance of plate-based controls for column 1 MXC (C1_MXC), column 2 MXC in metabolism negative mode (C2_MXC(−)), and column 2 MXC in metabolism positive mode (C2_MXC(+)). Plate-based normalization was calculated the same as E2 (D). Data points are plate-level mean normalized responses for all experimental plates (n=27).

For both modes of the metabolism assay, column 1 AIME lid pegs were left untreated (no alginate microsphere) to maintain standardized conditions described in the VM7Luc ERTA assay design. During screening, all control and test chemical raw data were normalized back to plate-level column 1 E2 controls to ensure consistent data normalization across the metabolism modes and assay conditions. Mean E2 efficacy values in metabolism negative mode was 87.6% of baseline control, indicating a marginal drop in E2 activity in the presence of the blank alginate microsphere (Figure 3C). In metabolism positive mode, mean E2 efficacy levels dropped to 56.3%, suggesting metabolism mediated partial inactivation of the positive control compound (Tsuchiya et al., 2005). For MXC, the mean value for metabolism negative controls also decreased relative to standard MXC conditions, but as expected, the mean efficacy increased in metabolism positive mode (Figure 3D).

Assay summary performance metrics were determined for inter-assay variability and screening quality across metabolism modes and cell culture medium conditions (Table 1). For inter-assay variability, the CV for the DMSO solvent controls in guideline conditions (metabolism negative, NRS negative) was 6.75%. Supplementation with NRS alone had a modest increase in the variability to 8.93%. The presence of alginate-S9 (metabolism positive) further increased the variability to 17.17%. For the assay positive control, E2 variability was low in guideline (2.77%) and metabolism negative (2.82%) conditions. The variability increased to 8.51% in metabolism positive mode, likely owing to the decrease in activity observed (Figure 3C). Variability decreased for MXC under control conditions and increased slightly postmetabolism. For screening quality, the Z′-factor calculated using DMSO and E2 controls was 0.90 for the guideline condition. The value did not change considerably when NRS was supplemented in the medium in metabolism negative mode (0.91), but did decrease (0.69) once the assay was run in metabolism positive mode. The decrease in E2 precision and drop in dynamic range likely contributed to the reduction in screening quality parameters (Figure 3C). However, the overall summary assay performance data reflected low inter-assay variability (CV < 20%) and suitable screening quality (Z′ > 0.5) between metabolism modes and medium conditions.

Table 1.

Summary Assay Performance Statistics for Screening Variability and Quality

		Metabolism
		Neg	Pos	Neg	Pos	Neg	Pos	Neg	Pos
NRS	Neg	0.90	NA	6.75	NA	2.77	NA	5.39	NA
	Pos	0.91	0.69	8.93	17.17	2.82	8.51	2.98	5.23
		Z’		CV: DMSO		CV: E2		CV: MXC

Inter-assay variability for negative (DMSO), positive (17β-estradiol; E2), and metabolism (methoxychlor; MXC) controls was determined by coefficient of variation (CV). ERTA screening quality denoted by Z′-factor (Z′) for negative and positive controls. Performance parameters were evaluated formetabolism negative (Neg) and positive (Pos) modes of the AIME assay ± supplementation of the cell culture medium with an NADPH regeneration system (NRS).

Abbreviation: NA, not applicable.

AIME—VM7Luc ERTA Reference Compound Screening

The VM7Luc ERTA assay was conducted in agonist mode ± metabolism with reference chemicals intended to validate the ERTA target assay effects when coupled to the AIME metabolism assay. Raw ERTA and cytotoxicity data were analyzed in the ToxCast Pipeline (TCPL) (Filer et al., 2017) and for curve-fit uncertainty quantification using ToxBoot (Watt and Judson, 2018). The uncertainty analysis provided a quantitative determination of hit probability (%hitc) in the TCPL curve fits within each metabolism condition (Supplementary Tables 7–14). The resulting ERTA sampling distribution data were evaluated for potency (AC₅₀), efficacy (Max Efficacy), and AUC (Figure 4A). The change in mean AUC (ΔAUC) from the ToxBoot data enabled ranking of test compound effects by shift magnitude with confidence intervals that enabled discrimination of metabolism-dependent effects (Figure 4B). Any change in binary hit calls from TCPL for ERTA (ΔHitc_ER) further enabled comparison of directional shifts between metabolism assay modes as they related to ERTA threshold cutoffs (Table 2).

Figure 4.

AIME-coupled ERTA positive reference compound screening. The confidence intervals derived from the uncertainty quantification analysis for ER agonist reference compounds were plotted for area under the fitted curve (AUC), active concentration at 50% (AC₅₀), and maximum effect levels (Max Efficacy %) for metabolism negative (control; blue) and metabolism positive (metabolism; red) assay modes. Box and whisker plots represent minimum, 25th percentile, median, 75th percentile, and maximum values (A). The ΔAUC and confidence interval derived from ToxBoot uncertainty quantification analysis are plotted with mean ± quantile of 3. Significance of metabolic effects is noted for bioinactivation (blue), no metabolic effect (black), and bioactivation (red) (B). ToxCast pipeline (TCPL) curve fits for metabolism negative (control; blue) and metabolism positive (metabolism; red) assay modes. Vertical lines represent the potency (AC₅₀) for respective assay modes. Horizontal line is the mean response for test guideline E2 controls (grey solid line). Plots are shown for all positive reference chemicals. Data points are normalized responses plotted for all experimental replicates (n=3) (C).

Six ER agonist reference compounds (genistein; bisphenol A; daidzein; coumestrol; 17α-ethinylestradiol; and ethylparaben) from OECD TG 455 (OECD, 2016) were included in the screening set. As expected, all 6 were flagged as TCPL active hits in metabolism negative mode based on efficacy (Table 2 and Supplementary Table 7), supporting the estrogenic effect of the parent compounds. Median AC₅₀ values were within 10-fold of the potency values reported (OECD, 2016) for the VM7Luc ERTA assay in the absence of metabolism (Supplementary Table 6). The exception was 17α-ethinylestradiol where the tested concentration range was not broad enough to capture a no observed effect level, confounding an accurate AC₅₀ determination (Figure 4C). Genistein and bisphenol A had ΔAUC values that trended toward bioactivation in metabolism positive mode. Conversely, daidzein, coumestrol, 17α-ethinylestradiol, and ethylparaben trended toward bioinactivation (Figure 4B and Supplementary Table 7). Cytotoxicity was noted for bisphenol A, 17α-ethinylestradiol, and genistein, but the active concentration at cutoff (ACC) values were higher than those for ERTA (Supplementary Figure 2A and Supplementary Table 11). Overall, CI range and ΔHitc_ER were insignificant and neutral, respectively, for all reference compounds except ethylparaben (Table 2 and Supplementary Table 7). The negative ΔAUC (−62.9), negative CI [−72.1, −53.74], and shift in ΔHitc_ER (−1) for ethylparaben supported a complete loss of efficacy by apparent bioinactivation (Table 2).

Two additional controls, resveratrol and the methoxychlor metabolite HPTE, were included in the positive reference set. Resveratrol, an ERTA antagonist negative control (OECD, 2016), exhibited agonist activity in both assay modes (Figs. 4A–C), consistent with previously reported estrogenic effects (Gehm et al., 1997). In addition, HPTE also demonstrated concentration-dependent effects that were fairly consistent between the metabolism modes (Figs. 4A–C). Resveratrol and HPTE were cytotoxic in one or both assay modes, but only resveratrol had ERTA bioactivity at or near the cytotoxic threshold (Supplementary Figure 2A and Supplementary Table 11). Neither compound demonstrated a significant metabolismdependent shift in activity (Figure 4B).

Three ER antagonist reference compounds ((Z)-4-hydroxytamoxifen; fulvestrant; and tamoxifen) and 3 agonist negative control chemicals (atrazine; corticosterone; and spironolactone) were also tested (Supplementary Figs. 1A–C). The activity of these compounds in ERTA agonist mode was expected to be primarily inactive. Tamoxifen and corticosterone were flagged as TCPL hit calls in metabolism negative mode (Table 2), but exhibited marginal bioactivity just above the threshold (Supplementary Figure 1C). Shifts in bioactivity between metabolism modes were not significant (Supplementary Figure 1B). Spironolactone, tamoxifen, and (Z)-4-hydroxytamoxifen were identified as cytotoxic (Supplementary Table 12). Spironolactone toxicity was not observed following metabolism, whereas activity persisted for the other 2 compounds (Supplementary Figure 3A and Supplementary Table 12). Tamoxifen ERTA potency (ACC) was near, or overlapped, with the cytotoxicity ACC in both assay modes (Supplementary Figure 3A). In summary, the data derived from the positive and negative reference chemicals supported high sensitivity and specificity for the VM7Luc ERTA assay that conformed to test guideline expectations when coupled to the AIME assay, deeming it suitable to evaluate test chemicals in both AIME metabolism modes.

AIME—VM7Luc ERTA Assay Metabolism Positive Test Compound Screening

The metabolism positive test set consisted of 34 compounds previously identified to produce metabolites that were more estrogenic than the parent compound (Pinto et al., 2016). Ranking of chemicals by ΔAUC yielded 19/34 compounds with metabolism-dependent activation shifts (positive ΔAUC) (Figure 5A and Supplementary Table 9). Of these, 11 had CI ranges > 0 (methoxychlor; trans-stilbene; chalcone; azobenzene; trans-α-methylstilbene; 1,1′-(1,2-ethynediyl)bis-benzene; 2-hydroxy-4-methoxybenzophenone; 2,2-diphenylpropane; 4-tert-butylphenyl salicylate; diphenylmethane; and biphenyl) demonstrating clear separation between parent and metabolite effects, and lending confidence to increased metabolite estrogenicity (Figure 5B). It is worth noting only 2/11 bioactivated compounds (diphenylmethane and biphenyl) exhibited transitions in TCPL hit calls. Each of the 11 compounds, with the exception of methoxychlor, also displayed a clear increase in max efficacy that was likely the major contribution to the ΔAUC shift (Supplementary Figure 6A). In contrast, the change in methoxychlor AUC was due to a more notable increase in potency, as well as a modest increase in efficacy (Supplementary Figure 6A).

Figure 5.

AIME-coupled ERTA metabolism positive test set screening. The ΔAUC and confidence interval derived from ToxBoot uncertainty quantification analysis are plotted with mean ± quantile of 3. Significance of metabolic effects is noted for bioinactivation (blue), no metabolic effect (black), and bioactivation (red) (A). ToxCast pipeline (TCPL) curve fits for metabolism negative (control; blue) and metabolism positive (metabolism; red) assay modes. Vertical lines represent the potency (AC₅₀) for respective assay modes. Horizontal line is the mean response for test guideline E2 controls (grey solid line). Plots are shown for chemicals where significant metabolism-dependent effects were observed. Data points are normalized responses plotted for all experimental replicates (n=3) (B).

Of the remaining compounds, 14/15 had inactivation shifts (negative ΔAUC) (Figure 5A and Supplementary Table 9). Only chrysene had a CI range < 0, indicating a metabolism-dependent effect, but neither the parent nor metabolite was active above the cutoff for ERTA (Figure 5B). TCPL hit calls were lost for 3 compounds (2-nitroflourene, fluorene, and 1,1′-(1,3-propanediyl)bis-benzene) indicating bioinactivation, though parent compounds were marginally active around the cutoff without metabolism (Supplementary file: ERTA_AEID2484 vs AEID2486_TCPL_Dual Plots). It is worth noting that the proestrogenic parent compounds formononetin, biochanin A, and mestranol all exhibited markedly high bioactivity. Despite expectations for bioactivation to more estrogenic metabolites (Schmider et al., 1997; Tolleson et al., 2002), activity decreased after metabolism, though the metabolic effect was not significant (Figure 5A).

Cytotoxicity was observed for 6 chemicals (biochanin A; benzophenone; phenol, 2-(phenylmethyl)-; 4-tert-butylphenyl salicylate; chalcone; and methoxychlor) in one or both assay modes (Supplementary Table 13), but ERTA ACC values for the corresponding active hit calls were well below the cytotoxicity range of most of these (Supplementary Figure 4A). Based on TCPL hit calls, metabolism-dependent cytotoxicity changes were observed for 4 chemicals (benzophenone; phenol, 2-(phenylmethyl)-; 4-tert-butylphenyl salicylate; and chalcone) (Supplementary Table 13). Of these, only chalcone and phenol, 2-(phenylmethyl)- had %hitc > 50%, suggesting bioactivated and bioinactivated cytotoxicity, respectively.

In conclusion, 11/34 compounds exhibited significant bioactivation in the positive test set within noncytotoxic ranges, supporting the value of the AIME platform in identifying bioactivated chemicals that may promote more significant estrogenicity following hepatic metabolism.

AIME—VM7Luc ERTA Assay Metabolism Negative Test Compound Screening

The metabolism negative test set was comprised of 14 compounds previously identified to have no or weak estrogenic activity of the parent molecules, and no evidence for conversion to estrogenic metabolites (Pinto et al., 2016). Ranking by ΔAUC resulted in 9 bioinactivation shifts (bumetrizole, 2-(2H-benzotriazol-2-yl)-4,6-bis(1,1-dimethylpropyl)phenol; 2-mercaptobenzothiazole; 2-(2H-benzotriazol-2-yl)-4-methylphenol; di-n-octyl phthalate; 2-ethylhexyl 4-(dimethylamino)benzoate; dicyclohexyl phthalate; dibutyl phthalate; and benzyl butyl phthalate), 4 neutral shifts due to absence of bioactivity (disulfiram; di(2-ethylhexyl) phthalate; bis(2-ethylhexyl)hexanedioate; and acrylonitrile), and one bioactivation shift (2,4-di-tert-butylphenol) (Figure 6A and Supplementary Table 10). Four bioinactivated compounds displayed significant metabolic effects (2-ethylhexyl 4-(dimethylamino)benzoate; dicyclohexyl phthalate; dibutyl phthalate; and benzyl butyl phthalate) indicating significant loss of ER activity with metabolism (Figure 6B). Two of these (2-ethylhexyl 4-(dimethylamino)benzoate and dicyclohexyl phthalate) did not show TCPL hit call shifts, highlighting the lapse in identification with TCPL cutoffs alone (Table 2). The bioactivated compound 2,4-di-tert-butylphenol had a significant gain in ER activity (Figs. 6A and 6B), and a gain in TCPL hit call (Supplementary Table 10), suggesting one bioactivated substance in the metabolism negative test set.

Figure 6.

AIME-coupled ERTA metabolism negative test set screening. The ΔAUC and confidence interval derived from ToxBoot uncertainty quantification analysis are plotted with mean ± quantile of 3. Significance of metabolic effects is noted for bioinactivation (blue), no metabolic effect (black), and bioactivation (red) (A). ToxCast pipeline (TCPL) curve fits for metabolism negative (control; blue) and metabolism positive (metabolism; red) assay modes. Vertical lines represent the potency (AC₅₀) for respective assay modes. Horizontal line is the mean response for test guideline E2 controls (grey solid line). Plots are shown for chemicals where significant metabolism-dependent effects were observed. Data points are normalized responses plotted for all experimental replicates (n=3) (B).

Of interest was of the complementary loss of cytotoxicity observed for 2,4-di-tert-butylphenol postmetabolism (Supplementary Table 14). No other cytotoxicity was evident in the metabolism negative test set. With the exception of 2,4-di-tert-butylphenol, the results for the negative test set were in agreement with weak evidence for metabolic conversion to estrogenic metabolites (Pinto et al., 2016). However, the loss of estrogenic activity observed for 2-ethylhexyl 4-(dimethylamino)benzoate, dicyclohexyl phthalate, dibutyl phthalate, and benzyl butyl phthalate following metabolism was less expected. The significant decrease in ERTA activity for these compounds emphasized the need to characterize the bioinactivation of estrogenic substances as well.

AIME—VM7Luc ERTA Assay: Relevance to the ToxCast ER Model and Uterotrophic Bioassay Data

The ToxCast ER computational model utilizes a set of 18 assays to calculate a weight of evidence score (AUC) for ER bioactive chemicals (Browne et al., 2015; Judson et al., 2015). The overall accuracy of the model (93%) is high and has good concordance (84%) with a set of reference chemicals evaluated against the “gold standard” uterotrophic bioassay (EPA US, 2009; OECD, 2007). To assess the utility of applying hepatic metabolism to the VM7Luc ERTA assay, results were compared with available data from the ToxCast ER model (Table 2). The in vitro assay results were also benchmarked against guideline-like uterotrophic (GL-UT) study data (Supplementary Table 15) (Kleinstreuer et al., 2016) in both metabolism modes to determine if metabolism modified concordance with the in vivo data. Of the 63 chemicals screened in this study, 43 had an associated score for agonist mode of the ToxCast ER model where 10 were active (0.1<AUC < 1), 15 were inconclusive (0<AUC < 0.1), and 18 were inactive (AUC = 0).

The parent compounds for all 10 active chemicals were flagged as positive TCPL hits (Hitc_Met_Neg), consistent with the ToxCast ER model scores for estrogen bioactivity (Table 2). A significant metabolic effect (Met_Effect) was only observed for benzyl butyl phthalate and methoxychlor. Benzyl butyl phthalate lost a TCPL hit call designation after metabolism (Hitc_Met_Pos) and was negative for uterotrophic weight-of-evidence (GL-WoE). The positive change in concordance (ΔMet) with in vivo findings also suggested improved in vivo prediction. In contrast, methoxychlor had no differential shift in TCPL hit calls (ΔHitc_ER), but demonstrated the highest bioactivation shift (ΔAUC) (Figure 5). The ToxCast ER model scores of 0.25 and 0.57 for methoxychlor and bis-demethylated metabolite HPTE, respectively, support higher in vitro estrogenic activity for the metabolite. In addition, the GL-WoE trends toward positive for methoxychlor. Therefore, the AIME-ERTA findings for methoxychlor are consistent with the in vivo results as a bioactivated substance.

Table 2.

Comparison of AIME-VM7Luc ERTA Metabolic Screening Results with the ToxCast ER Model and Rodent Uterotrophic Bioassay

			ToxCast ER Modela	Uterotrophic Studiesb			AIME - VM7Luc ERTAc						Concordance with In Vivod
CASRN	Chemical Name	Classification	AUC_Agonist	GL_Neg	GL_Pos	GL_WoE	Hitc_Met_Neg	Hitc_Met_Pos	ΔHitcER	ΔAUC	ΔAUC CI	Met_Effect	Met_Neg	Met_Pos	ΔMet
57–63–6	17α-Ethinylestradiol	Reference_Agonist	1.00	0	15	POS	1	1	0	−11.48	[−184.67, 161.71]	NEG	1	1	0
50–28–2	17β-Estradiol	Reference_Agonist	0.94	0	22	POS	1	1	0	NA	NA	NA	1	1	0
72–33–3	Mestranol	Metabolism_Positive	0.74	0	3	POS	1	1	0	−60.83	[−159.37, 37.7]	NEG	1	1	0
2971–36–0	2,2-Bis(4-hydroxyphenyl)-1, 1,1-trichloroethane	Reference_Agonist_ Methoxychlor metabolite HPTE	0.57	0	0	POS	1	1	0	3.95	[−27.16, 35.06]	NEG	NA	NA	NA
446–72–0	Genistein	Reference_Agonist	0.54	0	8	POS	1	1	0	27.96	[−1.37, 57.29]	NEG	1	1	0
80–05–7	Bisphenol A	Reference_Agonist	0.45	4	10	POS	1	1	0	1.57	[−46.01, 49.15]	NEG	1	1	0
486–66–8	Daidzein	Reference_Agonist	0.44	1	1	POS	1	1	0	−4.32	[−24.77, 16.12]	NEG	EQUIV	EQUIV	EQUIV
491–80–5	Biochanin A	Metabolism_Positive	0.36	0	0	POS	1	1	0	−38.41	[−166.17, 89.36]	NEG	NA	NA	NA
72–43–5	Methoxychlor	Metabolism_Positive	0.25	1	3	POS	1	1	0	83.56	[45.44,121.67]	POS	1	1	0
85–68–7	Benzyl butyl phthalate	Metaboli sm_Negative	0.18	1	0	POS	1	0	−1	−73.48	[−78.91, −68.05]	POS	0	1	1
2440–22–4	2-(2H-Benzotriazol-2-yl)-4-methylphenol	Metaboli sm_Negative	0.09	1	0	POS	1	1	0	−3.00	[−7.43, 1.43]	NEG	0	0	0
120–47–8	Ethylparaben	Reference_Agonist	0.09	2	1	POS	1	0	−1	−62.93	[−72.12, −53.74]	POS	0	1	1
131–57–7	2-Hydroxy-4-methoxybenzophenone	Metabolism_Positive	0.06	1	0	POS	1	1	0	39.65	[28.31, 50.98]	POS	0	0	0
87–18–3	4-tert-Butylphenyl salicylate	Metabolism_Positive	0.05	0	0	POS	1	1	0	23.39	[16.35, 30.42]	POS	NA	NA	NA
119–61–9	Benzophenone	Metabolism_Positive	0.04	1	1	POS	1	1	0	0.83	[−2.49, 4.14]	NEG	EQUIV	EQUIV	EQUIV
90–30–2	N-Phenyl-l-naphthylamine	Metabolism_Positive	0.04	0	0	POS	1	1	0	−4.78	[−13.93, 4.37]	NEG	NA	NA	NA
56–55–3	Benz(a) anthracene	Metabolism_Positive	0.04	1	1	POS	1	1	0	−2.60	[−13.66, 8.45]	NEG	EQUIV	EQUIV	EQUIV
84–74–2	Dibutyl phthalate	Metaboli sm_Negative	0.03	2	0	POS	1	0	−1	−12.16	[−14.39, −9.92]	POS	0	1	1
101–81–5	Diphenylmethane	Metabolism_Positive	0.02	0	0	POS	0	1	1	6.98	[4.84, 9.12]	POS	NA	NA	NA
10540–29–1	Tamoxifen	Reference_Antagonist	0.02	0	10	POS	1	1	0	−1.20	[−13.69, 11.29]	NEG	1	1	0
84–61–7	Dicyclohexyl phthalate	Metaboli sm_Negative	0.02	2	0	POS	1	1	0	−9.23	[−14.28, −4.18]	POS	0	0	0
52315–07–8	Cypermethrin	Metabolism_Positive	0.01	0	0	POS	1	1	0	−0.70	[−3.64, 2.25]	NEG	NA	NA	NA
21245–02–3	2-Ethylhexyl 4-(dimethylamino) benzoate	Metaboli sm_Negative	0.01	0	0	POS	1	1	0	−6.95	[−11.03, −2.87]	POS	NA	NA	NA
206–44–0	Fluoranthene	Metabolism_Positive	0.01	0	0	POS	1	1	0	−2.53	[−8.55, 3.5]	NEG	NA	NA	NA
90–43–7	2-Phenylphenol	Metabolism_Positive	0.01	0	0	POS	1	1	0	1.05	[−3.31, 5.41]	NEG	NA	NA	NA
52–01–7	Spironolactone	Reference_Agonist Negative	0	0	0	POS	0	0	0	−1.44	[−5.94, 3.05]	NEG	NA	NA	NA
50–22–6	Corticosterone	Reference_Agonist Negative	0	0	0	POS	1	0	−1	−0.21	[−9.65, 9.23]	NEG	NA	NA	NA
1912–24–9	Atrazine	Reference_Agonist Negative	0	1	0	POS	0	0	0	0	[0,0]	NA	1	1	0
129453–61–8	Fulvestrant	Reference_Antagonist	0	0	0	POS	0	0	0	0	[0,0]	NA	NA	NA	NA
149–30–4	2-Mercaptobenzothiazole	Metaboli sm_Negative	0	1	0	POS	0	0	0	−2.16	[−8.33, 4.01]	NEG	1	1	0
107–13–1	Acrylonitrile	Metaboli sm_Negative	0	0	0	POS	0	0	0	0	[0,0]	NA	NA	NA	NA
103–23–1	Bis(2-ethylhexyl)hexanedioate	Metaboli sm_Negative	0	2	0	POS	0	0	0	0	[0,0]	NA	1	1	0
117–81–7	Di(2-ethylhexyl) phthalate	Metaboli sm_Negative	0	2	0	POS	0	0	0	0	[0,0]	NA	1	1	0
117–84–0	Di-n-octyl phthalate	Metaboli sm_Negative	0	0	0	POS	1	1	0	−3.51	[−11.36,4.33]	NEG	NA	NA	NA
97–77–8	Disulfiram	Metaboli sm_Negative	0	1	0	POS	0	0	0	0	[0,0]	NA	1	1	0
96–76–4	2,4-Di-tert-butylphenol	Metaboli sm_Negative	0	1	0	POS	0	1	1	7.25	[4.5,10.01]	POS	1	0	−1
86–73–7	Fluorene	Metabolism_Positive	0	0	0	POS	1	0	−1	−0.78	[−9.22, 7.65]	NEG	NA	NA	NA
85–01–8	Phenanthrene	Metabolism_Positive	0	0	0	POS	1	0	−1	0.19	[−1.8, 2.18]	NEG	NA	NA	NA
103–33–3	Azobenzene	Metabolism_Positive	0	0	0	POS	1	1	0	51.09	[44.07, 58.1]	POS	NA	NA	NA
91–20–3	Naphthalene	Metabolism_Positive	0	0	0	POS	0	0	0	−4.09	[−12.23, 4.06]	NEG	NA	NA	NA
52645–53−1	Permethrin	Metabolism_Positive	0	1	1	POS	1	1	0	0.71	[−3.7, 5.13]	NEG	EQUIV	EQUIV	EQUIV
129–00–0	Pyrene	Metabolism_Positive	0	0	0	POS	1	1	0	0.12	[−3.72, 3.96]	NEG	NA	NA	NA
92–52–4	Biphenyl	Metabolism_Positive	0	0	0	POS	0	1	1	5.90	[3.61, 8.19]	POS	NA	NA	NA
479–13–0	Coumestrol	Reference_Agonist	NA	0	0	POS	1	1	0	−7.94	[−28.58,12.7]	NEG	NA	NA	NA
68047–06–3	(Z)-4-Hydroxy tamoxifen	Reference_Antagonist	NA	0	1	POS	0	0	0	0	[0,0]	NA	0	0	0
501–36–0	Resveratrol	Reference_Antagonist Negative	NA	0	0	POS	1	1	0	10.07	[−14.97, 35.12]	NEG	NA	NA	NA
25973–55–1	2-(2H-Benzotriazol-2-yl)-4,6-bis (l,l-dimethylpropyl)phenol	Metaboli sm_Negative	NA	0	0	POS	0	0	0	−1.14	[−5.67, 3.38]	NEG	NA	NA	NA
3896–11–5	Bumetrizole	Metaboli sm_Negative	NA	0	0	POS	0	1	1	−0.45	[−11.06, 10.17]	NEG	NA	NA	NA
1081–75–0	1,1'-(l,3-Propanediyl)bis-benzene	Metabolism_Positive	NA	0	0	POS	1	0	−1	−0.51	[−3.18, 2.16]	NEG	NA	NA	NA
607–57–8	2-Nitrofluorene	Metabolism_Positive	NA	0	0	POS	1	0	−1	−2.44	[−8.49, 3.62]	NEG	NA	NA	NA
778–22–3	2,2-Diphenylpropane	Metabolism_Positive	NA	0	0	POS	1	1	0	34.03	[14.73, 53.34]	POS	NA	NA	NA
2219–82–1	2-tert-Butyl-6-methylphenol	Metabolism_Positive	NA	0	0	POS	1	1	0	−0.44	[−2.03, 1.16]	NEG	NA	NA	NA
50–32–8	Benzo (a) pyrene	Metabolism_Positive	NA	0	0	POS	1	1	0	−2.90	[−8.17, 2.38]	NEG	NA	NA	NA
28994–41–4	Phenol, 2-(phenylmethyl)-	Metabolism_Positive	NA	0	0	POS	1	1	0	2.34	[−4.86, 9.53]	NEG	NA	NA	NA
103–29–7	Bibenzyl	Metabolism_Positive	NA	0	0	POS	1	1	0	1.31	[−2.26, 4.88]	NEG	NA	NA	NA
94–41–7	Chalcone	Metabolism_Positive	NA	0	0	POS	1	1	0	51.87	[46.45, 57.3]	POS	NA	NA	NA
218–01–9	Chrysene	Metabolism_Positive	NA	0	0	POS	0	0	0	−7.41	[−13.51, −1.32]	POS	NA	NA	NA
501–65-5	1,1′-(1,2-Ethynediyl)bis-benzene	Metabolism_Positive	NA	0	0	POS	1	1	0	40.72	[34.78, 46.66]	POS	NA	NA	NA
485–72–3	Formononetin	Metabolism_Positive	NA	0	0	POS	1	1	0	−32.13	[−66.44, 2.17]	NEG	NA	NA	NA
1896–62–4	Methyl trans-styryl ketone	Metabolism_Positive	NA	0	0	POS	0	0	0	0	[0,0]	NA	NA	NA	NA
833–81–8	trans-a-Methylstilbene	Metabolism_Positive	NA	0	0	POS	1	1	0	50.16	[43.59, 56.73]	POS	NA	NA	NA
103–30–0	trans-Stilbene	Metabolism_Positive	NA	0	0	POS	1	1	0	80.87	[66.65, 95.09]	POS	NA	NA	NA
1439–07–2	trans-Stilbene oxide	Metabolism_Positive	NA	0	0	POS	1	1	0	2.21	[−1.37, 5.79]	NEG	NA	NA	NA

The 63 chemicals screened in the AIME-VM7Luc ERTA assay are listed with corresponding CASRN, chemical name, and AIME-ERTA classification.

^aToxCast ER Model(Browne et al., 2015) scores for agonist mode (AUC_Agonist).

^bUterotrophic data derived from guideline-like (GL) studies in the curated uterotrophic database(Kleinstreuer et al., 2016) identify negative (GL_Neg), positive (GL_Pos), and weight of evidence (GL_WoE).

^cResults for binary TCPL hit calls in metabolism negative (Hitc_Met_Neg) and positive (Hitc_Met_Pos) modes. Metabolism-dependent changes in ERTA TCPL hit call designation (ΔHitc_ER), shift in mean AUC (ΔAUC), confidence interval (ΔAUC CI), and significant metabolic effect (Met_Effect) are noted.

^dAIME-VM7Luc ERTA concordance (1) or non-concordance (0) to in uiuo uterotrophic study data (GL_WoE) for metabolism negative (Met_Neg) and positive (Met_Pos) assay modes. The change in concordance (ΔMet) is noted as the difference in metabolism versus no metabolism with increased (1), no change (0), or decreased (−1) predictivity.

Abbreviations: EQUIV, equivalent; NA, data not available.

Of the 15 inconclusive chemicals, 14 parent chemicals were consistent with the ToxCast ER model, denoted by positive TCPL hits (Hitc_Met_Neg) (Table 2). Metabolism-dependent effects (Met_Effect) were evident for 7 of these chemicals where 3 were bioactivated (2-hydroxy-4-methoxybenzophenone; 4-tert-butylphenyl salicylate; and diphenylmethane) and 4 bioinactivated (ethylparaben; dibutyl phthalate; dicyclohexyl phthalate; and 2-ethylhexyl 4-(dimethylamino)benzoate).

With respect to bioactivation, 2-hydroxy-4-methoxybenzophenone had the greatest ΔAUC shift, but no change in TCPL hit calls, so in vitro results were not concordant with the single GL-UT study.

In addition, without uterotrophic data available, no in vivo concordance could be determined for diphenylmethane or 4-tert-butylphenyl salicylate, but metabolism-dependent bioactivation provides stronger evidence for both as agonists in the ToxCast ER model. Of the bioinactivated compounds, postmetabolism TCPL hit call designations changed for 2/4 substances (ethylparaben and dibutyl phthalate), collectively supporting bioinactivation because concordance to in vivo results improved for both (ΔMet). Despite no shift in TCPL hit calls, metabolism-dependent effects for dicyclohexyl phthalate points to deprioritization as a weak agonist because in vitro bioinactivation was consistent with negative uterotrophic findings. No uterotrophic data were available for 2-ethylhexyl 4-(dimethylamino)benzoate, but significant in vitro bioinactivation suggests deprioritization could also be considered.

For the 18 chemicals designated inactive in the ToxCast ER model (Table 2), 3 (2,4-di-tert-butylphenol; azobenzene; and biphenyl) had significant bioactivation effects (Met_Effect). 2,4-di-tert-butylphenol gained a TCPL hit call postmetabolism, but the result was inconsistent with the single negative GL-UT study. Biphenyl also gained a TCPL hit call, but in vivo concordance could not be determined. Azobenzene had no differential hit call shift, but a considerable gain in AUC was observed. Concordance with uterotrophic data could not be determined, but based on the inactive ToxCast ER score designation, azobenzene could be reprioritized as a bioactivated substance for ER bioactivity. It is worth noting that 11/18 parent chemicals (spironolactone; atrazine; fulvestrant; 2-mercaptobenzothiazole; acrylonitrile; bis(2-ethylhexyl)hexanedioate; di(2-ethylhexyl) phthalate; disulfiram; 2,4-di-tert-butylphenol; naphthalene; and biphenyl) in the inactive set were consistent with the ToxCast ER model and had negative hit calls for ERTA bioactivity in the absence of metabolism.

Due to the intentional enrichment of biotransformed, estrogen-active chemicals in the metabolism positive test set, there were a number of compounds with significant metabolism-dependent changes, but for which no ToxCast ER model or uterotrophic data were available. Significant bioactivation was observed for 5 chemicals (trans-stilbene; transalpha-methylstilbene; 1,1′-(1,2-ethynediyl)bis-benzene; chalcone; and 2,2-diphenylpropane), as noted in previous in vitro assays (Pinto et al., 2016). All 5 parent chemicals were positive in the TCPL analysis, but the increased metabolite activity would justify higher priority ranking than the parent compounds. Indeed, a non-GL-UT study has observed positive estrogenic effects for trans-stilbene with increased effect for hydroxylated metabolites trans-4-hydroxystilbene and trans-4,4′-dihydroxystilbene at equivalent doses (Sanoh et al., 2006). Hence, the AIME metabolic approach is not limited to identification of false-negative and false-positive ERTA effects, but supported reprioritization of weak active or data-poor estrogenic chemicals and, in several cases, improved concordance with in vivo uterotrophic studies. Overall, the comparative analysis of in vitro and in vivo data supports integration of hepatic metabolism into HTS assays to refine the prioritization and prediction of estrogenic and nonestrogenic chemicals.

DISCUSSION

One limitation often noted in the U.S. EPA ToxCast and Tox21 HTS programs is the limited assay coverage of xenobiotic metabolic pathways. The objective of this study was to address this concern by developing a method for retrofitting HTS toxicological assays with hepatic metabolic competence. The AIME platform consists of custom 96-well microplate lids containing solid supports attached to hydrogel encapsulated hepatic S9. The method utilizes rat hepatic S9 with an NRS cocktail that supports phase I xenobiotic metabolism and minimizes cytotoxicity in cell-based assays. Assay medium can be conditioned with chemical-derived metabolites to broadly evaluate biological activity in a variety of cell-free and cell-based assays. Here, deployment to an ERTA assay with a test set of compounds previously identified to generate ER-active metabolites (Pinto et al., 2016) demonstrated utility for identifying bioactivated and bioinactivated substances, refining prioritization based on metabolic effects, and enhancing concordance to observed in vivo effects.

The liver is considered the primary route for endogenous metabolism of xenobiotics. The dominant phase I metabolic pathway encompasses enzymes integral to the oxidation, reduction, and hydrolysis of chemicals (Stanley, 2017). Experimental in vitro systems for the study of hepatic metabolism often use postmitochondrial supernatant (S9) fractions to enrich for soluble and membrane-bound phase I enzymes. Here, phenobarbital/β-naphthoflavone-induced male Sprague Dawley rat hepatic S9 was employed to capture an array of cytochrome P450 activity. Metabolic transformation was evident (Figure 2) across a reference set of probe substrates recommended for in vitro study of human cytochrome P450s (eg, 1A2, 2D6, 2B6, 2C9, and 2E1) (Bjornsson et al., 2003). It is recognized that incorporation of rat S9 in the AIME platform presents a species-dependent cytochrome P450 profile that is not analogous to humans and, due to induction, metabolizes compounds to a magnitude that is likely greater than basal metabolic rates in rat or human. Incorporation of exogenous rat S9 into in vitro assays is routinely used in genotoxicity testing where the strengths and limitations have been investigated (Cox et al., 2016). Induced rat S9 may certainly overpredict metabolism-dependent effects, but is the most conservative approach to reduce potential false-negative or false-positive observations. Attempts to incorporate human S9 into the assay resulted in decreased signal from reference chemicals, consistent with expected toxicokinetic differences between rodent and human (Ozawa et al., 2000). In the case of the ERTA assay, human S9 also resulted in assay interference for ER transactivation (ie, false-positive signal), owing to possible presence of endogenous estrogenic substances (data not shown). To minimize functional variability due to human donor sourcing (Parkinson, 1996; Pearce et al., 1996), and maximize identification of potential metabolite effects for HTS, rat S9 was deemed the more suitable option. Despite the technical limitations observed in this study with human S9, this does not discount the desire for data derived from human metabolism systems to conduct comprehensive human risk assessments. Moreover, it is plausible that optimization of the current AIME design for phase I oxidation alone could generate reactive intermediates that may otherwise be short-lived due to phase II conjugation in vivo, resulting in false-positive effects. In addition, parent chemicals that yield functional metabolites strictly via phase II metabolism would also not be identified in the current AIME design largely due to the lack of appropriate cofactors. These limitations provide context to prudent interpretation of biological relevance and could be addressed, if desirable, by modifying and optimizing the design for alternative metabolic pathways.

Alginate/PLL entrapment has been used previously to create HTS-compatible systems that immobilize cells on plastic pillar inserts to evaluate the impact of drug and chemical exposures on toxicity (Lee et al., 2008; Yu et al., 2018). In the AIME assay, replacing cells with active metabolic enzymes in the alginate microsphere enabled greater flexibility to incorporate metabolism into a wider breadth of cell-based assays that might otherwise be incompatible with alginate encapsulation due to cell type or assay detection technology. In addition, S9 encapsulation improved application in cell-based assays because cytotoxic effects stemming from S9-generated microsomal lipid peroxides may have been minimized (Yamamoto et al., 2011). With regard to small molecule diffusion (Figure 2), the mean molecular weight of the reference substrates (252.1Da; range 169.6–370.3Da) was consistent with the mean molecular weight (249.6Da; range 42.0–1355.4Da) observed across the 6408 chemicals in the ToxCast inventory (https://comptox.epa.gov/dashboard/chemical_lists/EPACHEMINV_AVAIL; last accessed September 28, 2020). Depletion of parent substrates with concomitant accumulation of respective metabolites provided evidence for small molecule diffusion in and out of the microspheres (Figure 2), alleviating concerns about significant sequestration effects or chemical library compatibility.

The stability of hepatic S9 enzyme activity and metabolism incubation period were important considerations because metabolic rates vary by substrate (Figure 2). For enzyme stability, the S9 needed to maintain activity throughout assay setup and upon encapsulation in alginate microspheres. Reducing the S9/alginate ratio down to 10% (vol/vol) did not significantly reduce the metabolic activity of the S9 as compared with higher ratios (data not shown), and was thought to minimize potential cytotoxicity, reduce protein-binding effects (Heringa et al., 2004; Hoogenboom et al., 2002), and lower the cost of the assay. Temporal monitoring of enzyme kinetics out to 8 h revealed the majority of enzyme activity occurred within the first 2 h of incubation for the majority of cytochrome P450 probe substrates (Figure 2). Enzyme activity was sustained thereafter, but appeared to plateau for most reference chemicals by 8 h. Progressive loss of enzyme activity is a common feature of S9 fractions, with many commercial vendors running metabolic stability experiments for only 30–60 min. Microsome reactions performed beyond 2 h at 37°C are typically not advised because the potential for thermal degradation may outweigh the benefit of marginally increased metabolic activation (Pearce et al., 1996). For practical purposes, the metabolism period for the AIME assay was appropriate to the functional features, and likely suitable for most rapidly metabolized compounds, but could potentially overlook chemicals biotransformed at lower reaction velocities.

For general application to diverse chemical libraries, the concentration range and spacing of test compounds had to be compatible with both the AIME metabolism assay and the coupled ERTA assay (Figure 3). Target substrates and toxicokinetic variables, such as Michaelis-Menten constant (Km), reaction velocity (V_max), or intrinsic clearance rates (CL_int), are largely unknown for every conceivable enzyme by chemical pair in HTS experiments. Thus, the chemical concentration intervals should be enriched for spacing relative to an unknown K_m, increasing the number of observations at or above the K_m, and maximizing reaction velocities to observe changes in bioactivity. Hence, the selected range (0.002–200 μM) exceeded the top concentration typically evaluated in a ToxCast/Tox21 assay (100 μM) and contained interval spacing that was biased toward the higher test concentrations. This approach appeared to work well for the ERTA reference and test chemicals, providing sufficient data points postmetabolism for accurate curve fitting and data analysis.

Exogenous S9 fractions have previously been used to examine the impact of metabolism on estrogen-active substances (Charles et al., 2000; Mollergues et al., 2017; Taxvig et al., 2011; van Vugt-Lussenburg et al., 2018). Here, deployment of the AIME platform to the VM7Luc ERTA assay was evaluated using E2 and MXC controls in the context of VM7Luc test guideline (no alginate microsphere; no NRS), metabolism negative (alginate microsphere; NRS), and metabolism positive (S9-alginate microsphere; NRS) conditions (Figure 3). Strict monitoring of these conditions was important because each distinguishing variable (eg, NRS and S9) affected VM7Luc4E2 transactivation response to test chemicals (Table 1). Using the test guideline E2 response for plate-level data normalization enabled equivalent response efficacy between metabolism negative and positive modes across the assay plates. E2 response efficacy from control conditions decreased 12% and 44% in negative and positive modes, respectively (Figure 3C). Metabolism of endogenous estrogens occurs primarily through oxidation in the liver (Tsuchiya et al., 2005), so the metabolism-dependent decrease in E2 activity was expected and has been observed with exogenous S9 metabolism in other ER reporter assays (Charles et al., 2000; Elsby et al., 2001; van Vugt-Lussenburg et al., 2018). The reason for the decrease in metabolism negative mode is not clear, but could be due to mild cellular stress induced by the NRS or modest sequestration of E2 into the empty alginate microsphere. The proestrogenic control compound, MXC, also exhibited a decrease in efficacy when transitioned from test guideline to metabolism negative conditions, but exhibited an increase in efficacy from 29% to 64% of E2 controls in metabolism positive mode (Figure 3D). MXC is metabolized to the potent metabolite HPTE and is well known to exhibit bioactivation for ER agonism (Mollergues et al., 2017; Shelby et al., 1996; van Vugt-Lussenburg et al., 2018). Hence, the coupling of metabolizing systems to existing in vitro assays should be carefully monitored for target assay variability, as well as any biotransformation effects on reference control compounds because performance may not fully conform to test guideline expectations. Optimized experimental designs would ideally incorporate reference chemicals for the target assay that are not metabolic substrates, reference chemicals that have known biotransformation effects in the target assay, and assay condition-specific performance metrics that are suitable for monitoring the metabolism and target assay components independently.

Identification of potential false-negative and false-positive ERTA effects has been a primary motivation for incorporating xenobiotic metabolism into HTS assays (Jacobs et al., 2013; OECD, 2014). In this study, concentration-dependent changes in VM7Luc ERTA efficacy was a remarkable feature both in the positive reference control (Figure 4) and metabolism test sets (Figs. 5 and 6). The serial treatment design, where parent chemicals were first metabolized in the AIME assay and conditioned medium subsequently transferred to the ERTA assay plate, likely had a significant impact on distinguishing parent from metabolite effects; particularly for rapidly metabolized compounds where biotransformation was clearly evident. For instance, complete postmetabolism elimination of ER bioactivity for ethylparaben (Figure 4C), a reference agonist for in vitro ER assays (OECD, 2016), can be explained by rapid clearance with human or rat S9 (Taxvig et al., 2011) and has been observed previously, albeit to a lesser degree, as decreased potency in the ER Calux assay (van Vugt-Lussenburg et al., 2018). Likewise, benzyl butyl phthalate, dibutyl phthalate, dicyclohexyl phthalate, and 2-ethylhexyl 4-(dimethylamino)benzoate displayed substantial loss of activity. All 5 compounds were considered active (benzyl butyl phthalate) or inconclusive (ethylparaben, dibutyl phthalate, dicyclohexyl phthalate, and 2-ethylhexyl 4-(dimethylamino)benzoate) in the ToxCast ER model (Browne et al., 2015; Judson et al., 2015) (Table 2). The addition of in vitro hepatic metabolism improved the concordance of these test compounds with the negative uterotrophic assay data (benzyl butyl phthalate, ethylparaben, and dibutyl phthalate), or simply provided additional evidence to inconclusive ToxCast ER model scores (dicyclohexyl phthalate), suggesting these could be reclassified as false positives or deprioritized for in vitro ERTA activity, respectively.

The chemicals in the metabolism test sets (Pinto et al., 2016) included compounds collected from a broad literature search intended to develop a training set for ER-QSAR analyses of parent and predicted phase I metabolites. Each study used varying combinations of in vitro metabolizing systems (eg, hepatic S9 fractions, microsomes) derived from different species (eg, rat, human), varied application of postmetabolism mixtures (eg, conditioned medium, purified fractions), different concentration series and exposure periods, and a number of ERTA target assays with differential sensitivity (eg, yeast 2-hybrid, transiently transfected or stable cell lines in multiple tissue- and species-derived cell types) and largely unknown degrees of intrinsic metabolic competence. The lack of consistency across experimental designs means it was not possible to strictly compare estrogenic bioactivity because standards are currently not established for defined estrogenic metabolites. However, metabolic activation was a significant outcome in the metabolism test sets where 12 chemicals (methoxychlor; trans-stilbene; chalcone; azobenzene; trans-a-methylstilbene; 1,1′-(1,2-ethynediyl)bis-benzene; 2-hydroxy-4-methoxybenzophenone; 2,2-diphenylpropane; 4-tert-butylphenyl salicylate; diphenylmethane; biphenyl; and 2,4-di-tert-butylphenol) exhibited metabolites with increased ER activation (Figs. 5 and 6). Diphenylmethane, biphenyl, and 2,4-di-tert-butylphenol parent compounds were not identified as active via TCPL analysis, but these were the 3 weakest shifts in AUC with concentration-response activity just above cutoff threshold. More striking were the other 9 parent chemicals with weak to moderate ERTA bioactivity that yielded metabolites with activity near or exceeding the E2 steroidal control. A strict binary hit call scheme was deemed inadequate to appropriately prioritize these metabolism-dependent effects because shift in activity occurred outside the bounds of the assay cutoff. The implementation of uncertainty quantification analysis (Watt and Judson, 2018) for the TCPL curve-fit data to evaluate the magnitude in ΔAUC enabled statistical discrimination of these relative changes in ER activation, based on efficacy and potency, that could be used to rank-order chemicals for further evaluation in a cutoff-independent manner. Depending on classification schemes, it seems evident that several of these could be reclassified from inactive or weak, to moderate or strong ER agonists, if screened in additional estrogen assays with metabolism.

A number of studies have examined the impact of in vitro metabolism on proestrogenic substrates including certain pesticides (Elsby et al., 2001; Kitamura et al., 2003), stilbene derivatives (Ogawa et al., 2006; Sanoh et al., 2003), chalcone flavonoids (Kohno et al., 2005), and UV-filters (Watanabe et al., 2015). Most of these studies compare the bioactivity of a highly purified parent and predicted or known metabolite(s) to provide insight into the full range of ER transactivation that is theoretically possible for a given substance (Elsby et al., 2001; Kitamura et al., 2003; Kohno et al., 2005; Sanoh et al., 2003; Watanabe et al., 2015). Where possible, these data can serve as guidelines, particularly for reference compounds, for expectations of potency and efficacy when evaluated within the scope of in vitro metabolizing systems. However, subsequent interpretation of metabolic effects will be dependent on the active enzyme profile of the metabolizing system, reaction kinetics of the compound, the number and stability of metabolites produced, and the directional nature (activation or inactivation) of the metabolic effect. In this study, MXC and the bis-demethylated metabolite, HPTE, supported a parent-metabolite ERTA bioactivity range paradigm where the potency of MXC (AC₅₀: 4.72 μM) was nearly 100-fold less than HPTE (AC₅₀: 0.05 μM), without any remarkable changes noted in efficacy (Max Effect: 60.1% vs 63.3%) (Supplementary Table 16). After metabolism of MXC by S9, the potency (AC₅₀: 1.03 μM) had increased approximately 5-fold with similar pre-metabolism and postmetabolism efficacy (60.1% vs 52.7%) (Supplementary Table 16). The weaker shift in potency relative to HPTE likely reflects the incomplete metabolism and production of other metabolites from MXC (Elsby et al., 2001). Increased potency was indeed observed for 10/12 bioactivated compounds (methoxychlor; 2-hydroxy-4-methoxybenzophenone; 4-tert-butylphenyl salicylate; diphenylmethane; azobenzene; biphenyl; 2,2-diphenylpropane; chrysene; trans-α-methylstilbene; trans-stilbene), though effects were modest when contrasted to the changes in efficacy observed for most (Supplementary Table 16). Considerable gain of function in these cases could be attributed to increased abundance of metabolites with higher affinity for the ER, as well as additive effects of metabolites. It is important to note that the postmetabolism composition of any given substance in the AIME assay at each interval in the tested concentration range may reflect variable mixtures of unknown composition that could exert unique effects on ERTA. The presence or absence of active metabolites at each concentration of the parent compound determine a data curve fit that may or may not reflect the same profile observed for pure metabolites, and should serve as a point of caution for quantitatively interpreting the activity of metabolized test substances. From a hazard identification standpoint, the AIME assay is suitable for identifying compounds with metabolic-dependent effects, but further investigation would be necessary to identify the active metabolites, and related activity, if deemed essential for a risk assessment.

There were a total of 20/63 chemicals tested with GL-UT weight of evidence that could be characterized as positive (8 chemicals) or negative (12 chemicals) (Table 2; GL_WoE). Of the positive chemicals, 7/8 (88%) of the parent chemicals were predicted to be estrogenic in the absence of in vitro metabolism (Table 2; Met_Neg). Only methoxychlor was deemed to have a significant metabolic shift in the AIME assay (Table 2; Met_Effect), but this did not change the in vivo concordance (Table 2; ΔMet). The impact of incorporating in vitro metabolism was more prevalent for the GL-UT negative chemicals where 6/12 (50%) of the parent chemicals were initially concordant (Table 2; Met_Neg). A metabolic effect (Table 2; Met_Effect) was noted for 6/12 of the chemicals where 3 compounds demonstrated improved concordance, one lost concordance, and 2 had no change (Table 2; ΔMet); resulting in a net effect of 8/12 (67%) concordance (Table 2; Met_Pos). Benzyl butyl phthalate, dibutyl phthalate, and ethylparaben were the 3 compounds that resulted in significant inactivation to improve in vivo concordance. Benzyl butyl phthalate is extensively metabolized in vivo to 6 major metabolites including hippuric acid, and monoesters monobutyl phthalate and monobenzyl phthalate (Nativelle et al., 1999). Dibutyl phthalate is also metabolized primarily to monobutyl phthalate (Anderson et al., 2001; Koch et al., 2007). Dicyclohexyl phthalate metabolism trended toward inactivation, but was not sufficiently inactivated to diminish in vitro estrogenic activity. Monoester metabolites derived from benzyl butyl phthalate (monobutyl phthalate and monobenzyl phthalate) and dibutyl phthalate (monobutyl phthalate) are inactive according to ToxCast ER model scores (Judson et al., 2015). Ethylparaben, on the other hand, is hydrolyzed by phase I carboxylesterases to the 4-hydroxybenzoic acid metabolite (Ozaki et al., 2013). Like the phthalate monoesters, 4-hydroxybenzoic acid is also inactive in the ToxCast ER model (Judson et al., 2015). The observations for improved in vivo concordance for these phthalates and parabens were likely a result of rapid phase I hydrolysis to metabolites that were inactive in the ERTA assay. Overall, the results of the limited sampling across these 20 chemicals suggest in vitro metabolism made little difference in predicting estrogenic activity of parent chemicals, but improved identification of false-positive ERTA results with phase I-dependent biotransformation. Additional sampling will be necessary to evaluate the utility in identifying both false-positive and false-negative ERTA results when compared with in vivo data supported in GL-UT studies.

In conclusion, the AIME method is a viable approach to retrofitting HTS assays with hepatic metabolism. Deployment to an ERTA assay enabled evaluation of key design considerations for coupling metabolism to existing test guideline assays for endocrine activity, as well as other assays in the ToxCast/Tox21 portfolio. Importantly, the findings lend support to refining the prioritization and prediction of estrogenic and nonestrogenic chemicals using an in vitro approach with metabolic competence and may be a useful new approach method for future endocrine-related screening in the EDSP.