Machine Learning Models for Estrogen Receptor Bioactivity and Endocrine Disruption Prediction.

The U.S. Environmental Protection Agency (EPA) periodically releases in vitro data across a variety of targets, including the estrogen receptor (ER). In 2015, the EPA used these data to construct mathematical models of ER agonist and antagonist pathways to prioritize chemicals for endocrine disruption testing. However, mathematical models require in vitro data prior to predicting estrogenic activity, but machine learning methods are capable of prospective prediction from the molecular structure alone. The current study describes the generation and evaluation of Bayesian machine learning models grouped by the EPA's ER agonist pathway model using multiple data types with proprietary software, Assay Central. External predictions with three test sets of in vitro and in vivo reference chemicals with agonist activity classifications were compared to previous mathematical model publications. Training data sets were subjected to additional machine learning algorithms and compared with rank normalized scores of internal five-fold cross-validation statistics. External predictions were found to be comparable or superior to previous studies published by the EPA. When assessing six additional algorithms for the training data sets, Assay Central performed similarly at a reduced computational cost. This study demonstrates that machine learning can prioritize chemicals for future in vitro and in vivo testing of ER agonism.

Comparing Machine Learning Models for Aromatase (P450 19A1).

Aromatase, or cytochrome P450 19A1, catalyzes the aromatization of androgens to estrogens within the body. Changes in the activity of this enzyme can produce hormonal imbalances that can be detrimental to sexual and skeletal development. Inhibition of this enzyme can occur with drugs and natural products as well as environmental chemicals. Therefore, predicting potential endocrine disruption via exogenous chemicals requires that aromatase inhibition be considered in addition to androgen and estrogen pathway interference. Bayesian machine learning methods can be used for prospective prediction from the molecular structure without the need for experimental data. Herein, the generation and evaluation of multiple machine learning models utilizing different sources of aromatase inhibition data are described. These models are applied to two test sets for external validation with molecules relevant to drug discovery from the public domain. In addition, the performance of multiple machine learning algorithms was evaluated by comparing internal five-fold cross-validation statistics of the training data. These methods to predict aromatase inhibition from molecular structure, when used in concert with estrogen and androgen machine learning models, allow for a more holistic assessment of endocrine-disrupting potential of chemicals with limited empirical data and enable the reduction of the use of hazardous substances.

Comparison of Machine Learning Models for the Androgen Receptor.

The androgen receptor (AR) is a target of interest for endocrine disruption research, as altered signaling can affect normal reproductive and neurological development for generations. In an effort to prioritize compounds with alternative methodologies, the U.S. Environmental Protection Agency (EPA) used in vitro data from 11 assays to construct models of AR agonist and antagonist signaling pathways. While these EPA ToxCast AR models require in vitro data to assign a bioactivity score, Bayesian machine learning methods can be used for prospective prediction from molecule structure alone. This approach was applied to multiple types of data corresponding to the EPA's AR signaling pathway with proprietary software, Assay Central. The training performance of all machine learning models, including six other algorithms, was evaluated by internal 5-fold cross-validation statistics. Bayesian machine learning models were also evaluated with external predictions of reference chemicals to compare prediction accuracies to published results from the EPA. The machine learning model group selected for further studies of endocrine disruption consisted of continuous AC50 data from the February 2019 release of ToxCast/Tox21. These efforts demonstrate how machine learning can be used to predict AR-mediated bioactivity and can also be applied to other targets of endocrine disruption.

Quantitative Response-Response Relationships Linking Aromatase Inhibition to Decreased Fecundity are Conserved Across Three Fishes with Asynchronous Oocyte Development.

Quantitative adverse outcome pathways (qAOPs) describe quantitative response-response relationships that can predict the probability or severity of an adverse outcome for a given magnitude of chemical interaction with a molecular initiating event. However, the taxonomic domain of applicability for these predictions is largely untested. The present study began defining this applicability for a previously described qAOP for aromatase inhibition leading to decreased fecundity developed using data from fathead minnow (Pimephales promelas). This qAOP includes quantitative response-response relationships describing plasma 17ß-estradiol (E2) as a function of plasma fadrozole, plasma vitellogenin (VTG) as a function of plasma E2, and fecundity as a function of plasma VTG. These quantitative response-response relationships simulated plasma E2, plasma VTG, and fecundity measured in female zebrafish (Danio rerio) exposed to fadrozole for 21 days but not these responses measured in female Japanese medaka (Oryzias latipes). However, Japanese medaka had different basal levels of plasma E2, plasma VTG, and fecundity. Normalizing basal levels of each measurement to equal those of female fathead minnow enabled the relationships to accurately simulate plasma E2, plasma VTG, and fecundity measured in female Japanese medaka. This suggests that these quantitative response-response relationships are conserved across these three fishes when considering relative change rather than absolute measurements. The present study represents an early step toward defining the appropriate taxonomic domain of applicability and extending the regulatory applications of this qAOP.

Uptake and Disposition of Select Pharmaceuticals by Bluegill Exposed at Constant Concentrations in a Flow-Through Aquatic Exposure System.

The increasing use of pharmaceuticals has led to their subsequent input into and release from wastewater treatment plants, with corresponding discharge into surface waters that may subsequently exert adverse effects upon aquatic organisms. Although the distribution of pharmaceuticals in surface water has been extensively studied, the details of uptake, internal distribution, and kinetic processing of pharmaceuticals in exposed fish have received less attention. For this research, we investigated the uptake, disposition, and toxicokinetics of five pharmaceuticals (diclofenac, methocarbamol, rosuvastatin, sulfamethoxazole, and temazepam) in bluegill sunfish (Lepomis macrochirus) exposed to environmentally relevant concentrations (1000-4000 ng L-1) in a flow-through exposure system. Temazepam and methocarbamol were consistently detected in bluegill biological samples with the highest concentrations in bile of 4, 940, and 180 ng g-1, respectively, while sulfamethoxazole, diclofenac, and rosuvastatin were only infrequently detected. Over 30-day exposures, the relative magnitude of mean concentrations of temazepam and methocarbamol in biological samples generally followed the order: bile ≫ gut > liver and brain > muscle, plasma, and gill. Ranges of bioconcentration factors (BCFs) in different biological samples were 0.71-3960 and 0.13-48.6 for temazepam and methocarbamol, respectively. Log BCFs were statistically positively correlated to pH adjusted log Kow (that is, log Dow), with the strongest relations for liver and brain (r2 = 0.92 and 0.99, respectively), implying that bioconcentration patterns of ionizable pharmaceuticals depend on molecular status, that is, whether a pharmaceutical is un-ionized or ionized at ambient tissue pH. Methocarbamol and temazepam underwent rapid uptake and elimination in bluegill biological compartments with uptake rate constants (Ku) and elimination rate constants (Ke) at 0.0066-0.0330 h-1 and 0.0075-0.0384 h-1, respectively, and half-lives at 18.1-92.4 h. Exposure to mixtures of diclofenac, methocarbamol, sulfamethoxazole, and temazepam had little or no influence on the uptake and elimination rates, suggesting independent multiple uptake and disposition behaviors of pharmaceuticals by fish would occur when exposed to effluent-influenced surface waters.

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

Anthropogenic activities introduce complex mixtures into aquatic environments, necessitating mixture toxicity evaluation during risk assessment. There are many alternative approaches that can be used to complement traditional techniques for mixture assessment. Our study aimed to demonstrate how these approaches could be employed for mixture evaluation in a target watershed. Evaluations were carried out over 2 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 activity. Priority chemicals and mixtures were identified based on their relative contributions to estimated mixture pressure (based on cumulative toxic units) and via predictive assessments (random forest regression). Whole mixture assessments identified target sites for further evaluation including two sites targeted for industrial/urban chemical mixture effects assessment; three target sites for pharmaceutical mixture effects assessment; three target sites for further mixture characterization; and three low-priority sites. Analyses identified 14 mixtures and 16 chemicals that significantly contributed to cumulative effects, representing high or medium priority targets for further ecotoxicological evaluation, monitoring, or regulatory assessment. Overall, our study represents an important complement to single-chemical prioritizations, providing a comprehensive evaluation of the cumulative effects of mixtures detected in a target watershed. Furthermore, it demonstrates how different tools and techniques can be used to identify diverse facets of mixture risk and highlights strategies that can be considered in future complex mixture assessments. Environ Toxicol Chem 2023;42:1229-1256. © 2023 SETAC.

Case Study in 21st-Century Ecotoxicology: Using In Vitro Aromatase Inhibition Data to Predict Reproductive Outcomes in Fish In Vivo.

To reduce the use of intact animals for chemical safety testing, while ensuring protection of ecosystems and human health, there is a demand for new approach methodologies (NAMs) that provide relevant scientific information at a quality equivalent to or better than traditional approaches. The present case study examined whether bioactivity and associated potency measured in an in vitro screening assay for aromatase inhibition could be used together with an adverse outcome pathway (AOP) and mechanistically based computational models to predict previously uncharacterized in vivo effects. Model simulations were used to inform designs of 60-h and 10-21-day in vivo exposures of adult fathead minnows (Pimephales promelas) to three or four test concentrations of the in vitro aromatase inhibitor imazalil ranging from 0.12 to 260 µg/L water. Consistent with an AOP linking aromatase inhibition to reproductive impairment in fish, exposure to the fungicide resulted in significant reductions in ex vivo production of 17ß-estradiol (E2) by ovary tissue (≥165 µg imazalil/L), plasma E2 concentrations (≥74 µg imazalil/L), vitellogenin (Vtg) messenger RNA expression (≥165 µg imazalil/L), Vtg plasma concentrations (≥74 µg imazalil/L), uptake of Vtg into oocytes (≥260 µg imazalil/L), and overall reproductive output in terms of cumulative fecundity, number of spawning events, and eggs per spawning event (≥24 µg imazalil/L). Despite many potential sources of uncertainty in potency and efficacy estimates based on model simulations, observed magnitudes of apical effects were quite consistent with model predictions, and in vivo potency was within an order of magnitude of that predicted based on in vitro relative potency. Overall, our study suggests that NAMs and AOP-based approaches can support meaningful reduction and refinement of animal testing. Environ Toxicol Chem 2023;42:100-116. © 2022 SETAC. This article has been contributed to by U.S. Government employees and their work is in the public domain in the USA.

Case Study in 21st Century Ecotoxicology: Using In Vitro Aromatase Inhibition Data to Predict Short-Term In Vivo Responses in Adult Female Fish.

The present study evaluated whether in vitro measures of aromatase inhibition as inputs into a quantitative adverse outcome pathway (qAOP) construct could effectively predict in vivo effects on 17ß-estradiol (E2) and vitellogenin (VTG) concentrations in female fathead minnows. Five chemicals identified as aromatase inhibitors in mammalian-based ToxCast assays were screened for their ability to inhibit fathead minnow aromatase in vitro. Female fathead minnows were then exposed to 3 of those chemicals: letrozole, epoxiconazole, and imazalil in concentration-response (5 concentrations plus control) for 24 h. Consistent with AOP-based expectations, all 3 chemicals caused significant reductions in plasma E2 and hepatic VTG transcription. Characteristic compensatory upregulation of aromatase and follicle-stimulating hormone receptor (fshr) transcripts in the ovary were observed for letrozole but not for the other 2 compounds. Considering the overall patterns of concentration-response and temporal concordance among endpoints, data from the in vivo experiments strengthen confidence in the qualitative relationships outlined by the AOP. Quantitatively, the qAOP model provided predictions that fell within the standard error of measured data for letrozole but not for imazalil and epoxiconazole. However, the inclusion of measured plasma concentrations of the test chemicals as inputs improved model predictions, with all predictions falling within the range of measured values. Results highlight both the utility and limitations of the qAOP and its potential use in 21st century ecotoxicology. Environ Toxicol Chem 2021;40:1155-1170. © 2020 SETAC. This article has been contributed to by US Government employees and their work is in the public domain in the USA.

Adverse Outcome Pathway Network-Based Assessment of the Interactive Effects of an Androgen Receptor Agonist and an Aromatase Inhibitor on Fish Endocrine Function.

Predictive approaches to assessing the toxicity of contaminant mixtures have been largely limited to chemicals that exert effects through the same biological molecular initiating event. However, by understanding specific pathways through which chemicals exert effects, it may be possible to identify shared "downstream" nodes as the basis for forecasting interactive effects of chemicals with different molecular initiating events. Adverse outcome pathway (AOP) networks conceptually support this type of analysis. We assessed the utility of a simple AOP network for predicting the effects of mixtures of an aromatase inhibitor (fadrozole) and an androgen receptor agonist (17ß-trenbolone) on aspects of reproductive endocrine function in female fathead minnows. The fish were exposed to multiple concentrations of fadrozole and 17ß-trenbolone individually or in combination for 48 or 96 h. Effects on 2 shared nodes in the AOP network, plasma 17ß-estradiol (E2) concentration and vitellogenin (VTG) production (measured as hepatic vtg transcripts) responded as anticipated to fadrozole alone but were minimally impacted by 17ß-trenbolone alone. Overall, there were indications that 17ß-trenbolone enhanced decreases in E2 and vtg in fadrozole-exposed fish, as anticipated, but the results often were not statistically significant. Failure to consistently observe hypothesized interactions between fadrozole and 17ß-trenbolone could be due to several factors, including lack of impact of 17ß-trenbolone, inherent biological variability in the endpoints assessed, and/or an incomplete understanding of interactions (including feedback) between different pathways within the hypothalamic-pituitary-gonadal axis. Environ Toxicol Chem 2020;39:913-922. © 2020 SETAC.

De Facto Water Reuse: Bioassay suite approach delivers depth and breadth in endocrine active compound detection.

Although endocrine disrupting compounds have been detected in wastewater and surface waters worldwide using a variety of in vitro effects-based screening tools, e.g. bioassays, few have examined potential attenuation of environmental contaminants by both natural (sorption, degradation, etc.) and anthropogenic (water treatment practices) processes. This study used several bioassays and quantitative chemical analyses to assess residence-time weighted samples at six sites along a river in the northeastern United States beginning upstream from a wastewater treatment plant outfall and proceeding downstream along the stream reach to a drinking water treatment plant. Known steroidal estrogens were quantified and changes in signaling pathway molecular initiating events (activation of estrogen, androgen, glucocorticoid, peroxisome proliferator-activated, pregnane X receptor, and aryl hydrocarbon receptor signaling networks) were identified in water extracts. In initial multi-endpoint assays geographic and receptor-specific endocrine activity patterns in transcription factor signatures and nuclear receptor activation were discovered. In subsequent single endpoint receptor-specific bioassays, estrogen (16 of 18 samples; 0.01 to 28 ng estradiol equivalents [E2Eqs]/L) glucocorticoid (3 of 18 samples; 1.8 to 21 ng dexamethasone equivalents [DexEqs]/L), and androgen (2 of 18 samples; 0.95 to 2.1 ng dihydrotestosterone equivalents [DHTEqs]/L) receptor transcriptional activation occurred above respective assay method detection limits (0.04 ng E2Eqs/L, 1.2 ng DexEqs/L, and 0.77 ng DHTEqs/L) in multiple sampling events. Estrogen activity, the most often detected, correlated well with measured concentrations of known steroidal estrogens (r2 = 0.890). Overall, activity indicative of multiple types of endocrine active compounds was highest in wastewater effluent samples, while activity downstream was progressively lower, and negligible in unfinished treated drinking water. Not only was estrogenic and glucocorticoid activity confirmed in the effluent by utilizing multiple methods concurrently, but other activated signaling networks that historically received less attention (i.e. peroxisome proliferator-activated receptor) were also detected.