Early Drug-Induced Liver Injury Risk Screening: "Free," as Good as It Gets.

For all the promise of and need for clinical drug-induced liver injury (DILI) risk screening systems, demonstrating the predictive value of these systems versus readily available physicochemical properties and inherent dosing information has not been thoroughly evaluated. Therefore, we utilized a systematic approach to evaluate the predictive value of in vitro safety assays including bile salt export pump transporter inhibition and cytotoxicity in HepG2 and transformed human liver epithelial along with physicochemical properties. We also evaluated the predictive value of in vitro ADME assays including hepatic partition coefficient (Kp) and its unbound counterpart because they provide insight on hepatic accumulation potential. The datasets comprised of 569 marketed drugs with FDA DILIrank annotation (most vs less/none), dose and physicochemical information, 384 drugs with Kp and plasma protein binding data, and 279 drugs with safety assay data. For each dataset and combination of input parameters, we developed random forest machine learning models and measured model performance using the receiver operator characteristic area under the curve (ROC AUC). The median ROC AUC across the various data and parameters sets ranged from 0.67 to 0.77 with little evidence of additive predictivity when including safety or ADME assay data. Subsequent machine learning models consistently demonstrated daily dose, fraction sp3 or ionization, and cLogP/D inputs produced the best, simplest model for predicting clinical DILI risk with an ROC AUC of 0.75. This systematic framework should be used for future assay predictive value assessments and highlights the need for continued improvements to clinical DILI risk annotation.

Variability in in vivo studies: Defining the upper limit of performance for predictions of systemic effect levels.

New approach methodologies (NAMs) for chemical hazard assessment are often evaluated via comparison to animal studies; however, variability in animal study data limits NAM accuracy. The US EPA Toxicity Reference Database (ToxRefDB) enables consideration of variability in effect levels, including the lowest effect level (LEL) for a treatment-related effect and the lowest observable adverse effect level (LOAEL) defined by expert review, from subacute, subchronic, chronic, multi-generation reproductive, and developmental toxicity studies. The objectives of this work were to quantify the variance within systemic LEL and LOAEL values, defined as potency values for effects in adult or parental animals only, and to estimate the upper limit of NAM prediction accuracy. Multiple linear regression (MLR) and augmented cell means (ACM) models were used to quantify the total variance, and the fraction of variance in systemic LEL and LOAEL values explained by available study descriptors (e.g., administration route, study type). The MLR approach considered each study descriptor as an independent contributor to variance, whereas the ACM approach combined categorical descriptors into cells to define replicates. Using these approaches, total variance in systemic LEL and LOAEL values (in log10-mg/kg/day units) ranged from 0.74 to 0.92. Unexplained variance in LEL and LOAEL values, approximated by the residual mean square error (MSE), ranged from 0.20-0.39. Considering subchronic, chronic, or developmental study designs separately resulted in similar values. Based on the relationship between MSE and R-squared for goodness-of-fit, the maximal R-squared may approach 55 to 73% for a NAM-based predictive model of systemic toxicity using these data as reference. The root mean square error (RMSE) ranged from 0.47 to 0.63 log10-mg/kg/day, depending on dataset and regression approach, suggesting that a two-sided minimum prediction interval for systemic effect levels may have a width of 58 to 284-fold. These findings suggest quantitative considerations for building scientific confidence in NAM-based systemic toxicity predictions.

ToxRefDB version 2.0: Improved utility for predictive and retrospective toxicology analyses.

The Toxicity Reference Database (ToxRefDB) structures information from over 5000 in vivo toxicity studies, conducted largely to guidelines or specifications from the US Environmental Protection Agency and the National Toxicology Program, into a public resource for training and validation of predictive models. Herein, ToxRefDB version 2.0 (ToxRefDBv2) development is described. Endpoints were annotated (e.g. required, not required) according to guidelines for subacute, subchronic, chronic, developmental, and multigenerational reproductive designs, distinguishing negative responses from untested. Quantitative data were extracted, and dose-response modeling for nearly 28,000 datasets from nearly 400 endpoints using Benchmark Dose (BMD) Modeling Software were generated and stored. Implementation of controlled vocabulary improved data quality; standardization to guideline requirements and cross-referencing with United Medical Language System (UMLS) connects ToxRefDBv2 observations to vocabularies linked to UMLS, including PubMed medical subject headings. ToxRefDBv2 allows for increased connections to other resources and has greatly enhanced quantitative and qualitative utility for predictive toxicology.

Profiling 58 compounds including cosmetic-relevant chemicals using ToxRefDB and ToxCast.

Safety assessment for cosmetic-relevant chemicals (CRCs) in the European Union has been reshaped by restrictions on animal testing, and new approach methodologies (NAMs) for predicting toxicity are critical to ensure new cosmetic product safety. To demonstrate NAMs for safety assessment, we surveyed in vitro bioactivity and in vivo systemic toxicity data in the US Environmental Protection Agency's (EPA's) Toxicity Forecaster (ToxCast) and Toxicity Reference databases (ToxRefDB), respectively, for 58 chemicals identified as CRCs, including cosmetic ingredients as well as trace contaminants. CRCs were diverse in use types as suggested by broad chemical use categories. In terms of both target organ effects and study type, the median of the lowest effect level (LEL) doses in ToxRefDB for CRCs tended to be slightly higher than the median for the remaining 928 chemicals with study data in ToxRefDB, though the ranges of LELs were similar. For 17 of the 58 CRCs, high-throughput toxicokinetic data were used to calculate administered equivalent doses (AEDs) in mg/kg/day units for the in vitro bioactivity observed, and these AEDs served as conservative estimators of the systemic LELs observed in vivo. This work suggests that NAMs for bioactivity may inform a conservative point-of-departure estimate for diverse CRCs.

Novel application of normalized pointwise mutual information (NPMI) to mine biomedical literature for gene sets associated with disease: use case in breast carcinogenesis.

Advances in technology within biomedical sciences have led to an inundation of data across many fields, raising new challenges in how best to integrate and analyze these resources. For example, rapid chemical screening programs like the US Environmental Protection Agency's ToxCast and the collaborative effort, Tox21, have produced massive amounts of information on putative chemical mechanisms where assay targets are identified as genes; however, systematically linking these hypothesized mechanisms with in vivo toxicity endpoints like disease outcomes remains problematic. Herein we present a novel use of normalized pointwise mutual information (NPMI) to mine biomedical literature for gene associations with biological concepts as represented by Medical Subject Headings (MeSH terms) in PubMed. Resources that tag genes to articles were integrated, then cross-species orthologs were identified using UniRef50 clusters. MeSH term frequency was normalized to reflect the MeSH tree structure, and then the resulting GeneID-MeSH associations were ranked using NPMI. The resulting network, called Entity MeSH Co-occurrence Network (EMCON), is a scalable resource for the identification and ranking of genes for a given topic of interest. The utility of EMCON was evaluated with the use case of breast carcinogenesis. Topics relevant to breast carcinogenesis were used to query EMCON and retrieve genes important to each topic. A breast cancer gene set was compiled through expert literature review (ELR) to assess performance of the search results. We found that the results from EMCON ranked the breast cancer genes from ELR higher than randomly selected genes with a recall of 0.98. Precision of the top five genes for selected topics was calculated as 0.87. This work demonstrates that EMCON can be used to link in vitro results to possible biological outcomes, thus aiding in generation of testable hypotheses for furthering understanding of biological function and the contribution of chemical exposures to disease.

Screening the ToxCast phase II libraries for alterations in network function using cortical neurons grown on multi-well microelectrode array (mwMEA) plates.

Methods are needed for rapid screening of environmental compounds for neurotoxicity, particularly ones that assess function. To demonstrate the utility of microelectrode array (MEA)-based approaches as a rapid neurotoxicity screening tool, 1055 chemicals from EPA's phase II ToxCast library were evaluated for effects on neural function and cell health. Primary cortical networks were grown on multi-well microelectrode array (mwMEA) plates. On day in vitro 13, baseline activity (40 min) was recorded prior to exposure to each compound (40 µM). Changes in spontaneous network activity [mean firing rate (MFR)] and cell viability (lactate dehydrogenase and CellTiter Blue) were assessed within the same well following compound exposure. Following exposure, 326 compounds altered (increased or decreased) normalized MFR beyond hit thresholds based on 2× the median absolute deviation of DMSO-treated wells. Pharmaceuticals, pesticides, fungicides, chemical intermediates, and herbicides accounted for 86% of the hits. Further, changes in MFR occurred in the absence of cytotoxicity, as only eight compounds decreased cell viability. ToxPrint chemotype analysis identified several structural domains (e.g., biphenyls and alkyl phenols) significantly enriched with MEA actives relative to the total test set. The top 5 enriched ToxPrint chemotypes were represented in 26% of the MEA hits, whereas the top 11 ToxPrints were represented in 34% of MEA hits. These results demonstrate that large-scale functional screening using neural networks on MEAs can fill a critical gap in assessment of neurotoxicity potential in ToxCast assay results. Further, a data-mining approach identified ToxPrint chemotypes enriched in the MEA-hit subset, which define initial structure-activity relationship inferences, establish potential mechanistic associations to other ToxCast assay endpoints, and provide working hypotheses for future studies.

Use of high-throughput in vitro toxicity screening data in cancer hazard evaluations by IARC Monograph Working Groups.

Evidence regarding carcinogenic mechanisms serves a critical role in International Agency for Research on Cancer (IARC) Monograph evaluations. Three recent IARC Working Groups pioneered inclusion of the US Environmental Protection Agency (EPA) ToxCast program high-throughput screening (HTS) data to supplement other mechanistic evidence. In Monograph V110, HTS profiles were compared between perfluorooctanoic acid (PFOA) and prototypical activators across multiple nuclear receptors. For Monograph V112-113, HTS assays were mapped to 10 key characteristics of carcinogens identified by an IARC expert group, and systematically considered as an additional mechanistic data stream. Both individual assay results and ToxPi-based rankings informed mechanistic evaluations. Activation of multiple nuclear receptors in HTS assays showed that PFOA targets not only peroxisome proliferator activated receptors, but also other receptors. ToxCast assays substantially covered 5 of 10 key characteristics, corroborating literature evidence of "induces oxidative stress" and "alters cell proliferation, cell death or nutrient supply" and filling gaps for "modulates receptor-mediated effects." Thus, ToxCast HTS data were useful both in evaluating specific mechanistic hypotheses and in contributing to the overall evaluation of mechanistic evidence. However, additional HTS assays are needed to provide more comprehensive coverage of the 10 key characteristics of carcinogens that form the basis of current IARC mechanistic evaluations.

High-Throughput H295R Steroidogenesis Assay: Utility as an Alternative and a Statistical Approach to Characterize Effects on Steroidogenesis.

The U.S. Environmental Protection Agency Endocrine Disruptor Screening Program and the Organization for Economic Co-operation and Development (OECD) have used the human adrenocarcinoma (H295R) cell-based assay to predict chemical perturbation of androgen and estrogen production. Recently, a high-throughput H295R (HT-H295R) assay was developed as part of the ToxCast program that includes measurement of 11 hormones, including progestagens, corticosteroids, androgens, and estrogens. To date, 2012 chemicals have been screened at 1 concentration; of these, 656 chemicals have been screened in concentration-response. The objectives of this work were to: (1) develop an integrated analysis of chemical-mediated effects on steroidogenesis in the HT-H295R assay and (2) evaluate whether the HT-H295R assay predicts estrogen and androgen production specifically via comparison with the OECD-validated H295R assay. To support application of HT-H295R assay data to weight-of-evidence and prioritization tasks, a single numeric value based on Mahalanobis distances was computed for 654 chemicals to indicate the magnitude of effects on the synthesis of 11 hormones. The maximum mean Mahalanobis distance (maxmMd) values were high for strong modulators (prochloraz, mifepristone) and lower for moderate modulators (atrazine, molinate). Twenty-five of 28 reference chemicals used for OECD validation were screened in the HT-H295R assay, and produced qualitatively similar results, with accuracies of 0.90/0.75 and 0.81/0.91 for increased/decreased testosterone and estradiol production, respectively. The HT-H295R assay provides robust information regarding estrogen and androgen production, as well as additional hormones. The maxmMd from this integrated analysis may provide a data-driven approach to prioritizing lists of chemicals for putative effects on steroidogenesis.

An "EAR" on Environmental Surveillance and Monitoring: A Case Study on the Use of Exposure-Activity Ratios (EARs) to Prioritize Sites, Chemicals, and Bioactivities of Concern in Great Lakes Waters.

Current environmental monitoring approaches focus primarily on chemical occurrence. However, based on concentration alone, it can be difficult to identify which compounds may be of toxicological concern and should be prioritized for further monitoring, in-depth testing, or management. This can be problematic because toxicological characterization is lacking for many emerging contaminants. New sources of high-throughput screening (HTS) data, such as the ToxCast database, which contains information for over 9000 compounds screened through up to 1100 bioassays, are now available. Integrated analysis of chemical occurrence data with HTS data offers new opportunities to prioritize chemicals, sites, or biological effects for further investigation based on concentrations detected in the environment linked to relative potencies in pathway-based bioassays. As a case study, chemical occurrence data from a 2012 study in the Great Lakes Basin along with the ToxCast effects database were used to calculate exposure-activity ratios (EARs) as a prioritization tool. Technical considerations of data processing and use of the ToxCast database are presented and discussed. EAR prioritization identified multiple sites, biological pathways, and chemicals that warrant further investigation. Prioritized bioactivities from the EAR analysis were linked to discrete adverse outcome pathways to identify potential adverse outcomes and biomarkers for use in subsequent monitoring efforts.

Retrospective mining of toxicology data to discover multispecies and chemical class effects: Anemia as a case study.

Predictive toxicity models rely on large amounts of accurate in vivo data. Here, we analyze the quality of in vivo data from the U.S. EPA Toxicity Reference Database (ToxRefDB), using chemical-induced anemia as an example. Considerations include variation in experimental conditions, changes in terminology over time, distinguishing negative from missing results, observer and diagnostic bias, and data transcription errors. Within ToxRefDB, we use hematological data on 658 chemicals tested in one or more of 1738 studies (subchronic rat or chronic rat, mouse, or dog). Anemia was reported most frequently in the rat subchronic studies, followed by chronic studies in dog, rat, and then mouse. Concordance between studies for a positive finding of anemia (same chemical, different laboratories) ranged from 90% (rat subchronic predicting rat chronic) to 40% (mouse chronic predicting rat chronic). Concordance increased with manual curation by 20% on average. We identified 49 chemicals that showed an anemia phenotype in at least two species. These included 14 aniline moiety-containing compounds that were further analyzed for their potential to be metabolically transformed into substituted anilines, which are known anemia-causing chemicals. This analysis should help inform future use of in vivo databases for model development.

tcpl: the ToxCast pipeline for high-throughput screening data.

Motivation: Large high-throughput screening (HTS) efforts are widely used in drug development and chemical toxicity screening. Wide use and integration of these data can benefit from an efficient, transparent and reproducible data pipeline. Summary: The tcpl R package and its associated MySQL database provide a generalized platform for efficiently storing, normalizing and dose-response modeling of large high-throughput and high-content chemical screening data. The novel dose-response modeling algorithm has been tested against millions of diverse dose-response series, and robustly fits data with outliers and cytotoxicity-related signal loss. Availability and Implementation: tcpl is freely available on the Comprehensive R Archive Network under the GPL-2 license. Contact: martin.matt@epa.gov.

ToxCast Chemical Landscape: Paving the Road to 21st Century Toxicology.

The U.S. Environmental Protection Agency's (EPA) ToxCast program is testing a large library of Agency-relevant chemicals using in vitro high-throughput screening (HTS) approaches to support the development of improved toxicity prediction models. Launched in 2007, Phase I of the program screened 310 chemicals, mostly pesticides, across hundreds of ToxCast assay end points. In Phase II, the ToxCast library was expanded to 1878 chemicals, culminating in the public release of screening data at the end of 2013. Subsequent expansion in Phase III has resulted in more than 3800 chemicals actively undergoing ToxCast screening, 96% of which are also being screened in the multi-Agency Tox21 project. The chemical library unpinning these efforts plays a central role in defining the scope and potential application of ToxCast HTS results. The history of the phased construction of EPA's ToxCast library is reviewed, followed by a survey of the library contents from several different vantage points. CAS Registry Numbers are used to assess ToxCast library coverage of important toxicity, regulatory, and exposure inventories. Structure-based representations of ToxCast chemicals are then used to compute physicochemical properties, substructural features, and structural alerts for toxicity and biotransformation. Cheminformatics approaches using these varied representations are applied to defining the boundaries of HTS testability, evaluating chemical diversity, and comparing the ToxCast library to potential target application inventories, such as used in EPA's Endocrine Disruption Screening Program (EDSP). Through several examples, the ToxCast chemical library is demonstrated to provide comprehensive coverage of the knowledge domains and target inventories of potential interest to EPA. Furthermore, the varied representations and approaches presented here define local chemistry domains potentially worthy of further investigation (e.g., not currently covered in the testing library or defined by toxicity "alerts") to strategically support data mining and predictive toxicology modeling moving forward.

Evaluation of food-relevant chemicals in the ToxCast high-throughput screening program.

Thousands of chemicals are directly added to or come in contact with food, many of which have undergone little to no toxicological evaluation. The landscape of the food-relevant chemical universe was evaluated using cheminformatics, and subsequently the bioactivity of food-relevant chemicals across the publicly available ToxCast highthroughput screening program was assessed. In total, 8659 food-relevant chemicals were compiled including direct food additives, food contact substances, and pesticides. Of these food-relevant chemicals, 4719 had curated structure definition files amenable to defining chemical fingerprints, which were used to cluster chemicals using a selforganizing map approach. Pesticides, and direct food additives clustered apart from one another with food contact substances generally in between, supporting that these categories not only reflect different uses but also distinct chemistries. Subsequently, 1530 food-relevant chemicals were identified in ToxCast comprising 616 direct food additives, 371 food contact substances, and 543 pesticides. Bioactivity across ToxCast was filtered for cytotoxicity to identify selective chemical effects. Initiating analyses from strictly chemical-based methodology or bioactivity/cytotoxicity-driven evaluation presents unbiased approaches for prioritizing chemicals. Although bioactivity in vitro is not necessarily predictive of adverse effects in vivo, these data provide insight into chemical properties and cellular targets through which foodrelevant chemicals elicit bioactivity.

High-Throughput Screening of Chemical Effects on Steroidogenesis Using H295R Human Adrenocortical Carcinoma Cells.

Disruption of steroidogenesis by environmental chemicals can result in altered hormone levels causing adverse reproductive and developmental effects. A high-throughput assay using H295R human adrenocortical carcinoma cells was used to evaluate the effect of 2060 chemical samples on steroidogenesis via high-performance liquid chromatography followed by tandem mass spectrometry quantification of 10 steroid hormones, including progestagens, glucocorticoids, androgens, and estrogens. The study employed a 3 stage screening strategy. The first stage established the maximum tolerated concentration (MTC; ≥ 70% viability) per sample. The second stage quantified changes in hormone levels at the MTC whereas the third stage performed concentration-response (CR) on a subset of samples. At all stages, cells were prestimulated with 10 µM forskolin for 48 h to induce steroidogenesis followed by chemical treatment for 48 h. Of the 2060 chemical samples evaluated, 524 samples were selected for 6-point CR screening, based in part on significantly altering at least 4 hormones at the MTC. CR screening identified 232 chemical samples with concentration-dependent effects on 17ß-estradiol and/or testosterone, with 411 chemical samples showing an effect on at least one hormone across the steroidogenesis pathway. Clustering of the concentration-dependent chemical-mediated steroid hormone effects grouped chemical samples into 5 distinct profiles generally representing putative mechanisms of action, including CYP17A1 and HSD3B inhibition. A distinct pattern was observed between imidazole and triazole fungicides suggesting potentially distinct mechanisms of action. From a chemical testing and prioritization perspective, this assay platform provides a robust model for high-throughput screening of chemicals for effects on steroidogenesis.

Using ToxCast™ Data to Reconstruct Dynamic Cell State Trajectories and Estimate Toxicological Points of Departure.

BACKGROUND: High-content imaging (HCI) allows simultaneous measurement of multiple cellular phenotypic changes and is an important tool for evaluating the biological activity of chemicals. OBJECTIVES: Our goal was to analyze dynamic cellular changes using HCI to identify the "tipping point" at which the cells did not show recovery towards a normal phenotypic state. METHODS: HCI was used to evaluate the effects of 967 chemicals (in concentrations ranging from 0.4 to 200 µM) on HepG2 cells over a 72-hr exposure period. The HCI end points included p53, c-Jun, histone H2A.x, α-tubulin, histone H3, alpha tubulin, mitochondrial membrane potential, mitochondrial mass, cell cycle arrest, nuclear size, and cell number. A computational model was developed to interpret HCI responses as cell-state trajectories. RESULTS: Analysis of cell-state trajectories showed that 336 chemicals produced tipping points and that HepG2 cells were resilient to the effects of 334 chemicals up to the highest concentration (200 µM) and duration (72 hr) tested. Tipping points were identified as concentration-dependent transitions in system recovery, and the corresponding critical concentrations were generally between 5 and 15 times (25th and 75th percentiles, respectively) lower than the concentration that produced any significant effect on HepG2 cells. The remaining 297 chemicals require more data before they can be placed in either of these categories. CONCLUSIONS: These findings show the utility of HCI data for reconstructing cell state trajectories and provide insight into the adaptation and resilience of in vitro cellular systems based on tipping points. Cellular tipping points could be used to define a point of departure for risk-based prioritization of environmental chemicals. CITATION: Shah I, Setzer RW, Jack J, Houck KA, Judson RS, Knudsen TB, Liu J, Martin MT, Reif DM, Richard AM, Thomas RS, Crofton KM, Dix DJ, Kavlock RJ. 2016. Using ToxCast™ data to reconstruct dynamic cell state trajectories and estimate toxicological points of departure. Environ Health Perspect 124:910-919; http://dx.doi.org/10.1289/ehp.1409029.

Systems Toxicology of Male Reproductive Development: Profiling 774 Chemicals for Molecular Targets and Adverse Outcomes.

BACKGROUND: Trends in male reproductive health have been reported for increased rates of testicular germ cell tumors, low semen quality, cryptorchidism, and hypospadias, which have been associated with prenatal environmental chemical exposure based on human and animal studies. OBJECTIVE: In the present study we aimed to identify significant correlations between environmental chemicals, molecular targets, and adverse outcomes across a broad chemical landscape with emphasis on developmental toxicity of the male reproductive system. METHODS: We used U.S. EPA's animal study database (ToxRefDB) and a comprehensive literature analysis to identify 774 chemicals that have been evaluated for adverse effects on male reproductive parameters, and then used U.S. EPA's in vitro high-throughput screening (HTS) database (ToxCastDB) to profile their bioactivity across approximately 800 molecular and cellular features. RESULTS: A phenotypic hierarchy of testicular atrophy, sperm effects, tumors, and malformations, a composite resembling the human testicular dysgenesis syndrome (TDS) hypothesis, was observed in 281 chemicals. A subset of 54 chemicals with male developmental consequences had in vitro bioactivity on molecular targets that could be condensed into 156 gene annotations in a bipartite network. CONCLUSION: Computational modeling of available in vivo and in vitro data for chemicals that produce adverse effects on male reproductive end points revealed a phenotypic hierarchy across animal studies consistent with the human TDS hypothesis. We confirmed the known role of estrogen and androgen signaling pathways in rodent TDS, and importantly, broadened the list of molecular targets to include retinoic acid signaling, vascular remodeling proteins, G-protein coupled receptors (GPCRs), and cytochrome P450s. CITATION: Leung MC, Phuong J, Baker NC, Sipes NS, Klinefelter GR, Martin MT, McLaurin KW, Setzer RW, Darney SP, Judson RS, Knudsen TB. 2016. Systems toxicology of male reproductive development: profiling 774 chemicals for molecular targets and adverse outcomes. Environ Health Perspect 124:1050-1061; http://dx.doi.org/10.1289/ehp.1510385.

Integrated Model of Chemical Perturbations of a Biological Pathway Using 18 In Vitro High-Throughput Screening Assays for the Estrogen Receptor.

We demonstrate a computational network model that integrates 18 in vitro, high-throughput screening assays measuring estrogen receptor (ER) binding, dimerization, chromatin binding, transcriptional activation, and ER-dependent cell proliferation. The network model uses activity patterns across the in vitro assays to predict whether a chemical is an ER agonist or antagonist, or is otherwise influencing the assays through a manner dependent on the physics and chemistry of the technology platform ("assay interference"). The method is applied to a library of 1812 commercial and environmental chemicals, including 45 ER positive and negative reference chemicals. Among the reference chemicals, the network model correctly identified the agonists and antagonists with the exception of very weak compounds whose activity was outside the concentration range tested. The model agonist score also correlated with the expected potency class of the active reference chemicals. Of the 1812 chemicals evaluated, 111 (6.1%) were predicted to be strongly ER active in agonist or antagonist mode. This dataset and model were also used to begin a systematic investigation of assay interference. The most prominent cause of false-positive activity (activity in an assay that is likely not due to interaction of the chemical with ER) is cytotoxicity. The model provides the ability to prioritize a large set of important environmental chemicals with human exposure potential for additional in vivo endocrine testing. Finally, this model is generalizable to any molecular pathway for which there are multiple upstream and downstream assays available.

Predicting hepatotoxicity using ToxCast in vitro bioactivity and chemical structure.

The U.S. Tox21 and EPA ToxCast program screen thousands of environmental chemicals for bioactivity using hundreds of high-throughput in vitro assays to build predictive models of toxicity. We represented chemicals based on bioactivity and chemical structure descriptors, then used supervised machine learning to predict in vivo hepatotoxic effects. A set of 677 chemicals was represented by 711 in vitro bioactivity descriptors (from ToxCast assays), 4,376 chemical structure descriptors (from QikProp, OpenBabel, PaDEL, and PubChem), and three hepatotoxicity categories (from animal studies). Hepatotoxicants were defined by rat liver histopathology observed after chronic chemical testing and grouped into hypertrophy (161), injury (101) and proliferative lesions (99). Classifiers were built using six machine learning algorithms: linear discriminant analysis (LDA), Naïve Bayes (NB), support vector machines (SVM), classification and regression trees (CART), k-nearest neighbors (KNN), and an ensemble of these classifiers (ENSMB). Classifiers of hepatotoxicity were built using chemical structure descriptors, ToxCast bioactivity descriptors, and hybrid descriptors. Predictive performance was evaluated using 10-fold cross-validation testing and in-loop, filter-based, feature subset selection. Hybrid classifiers had the best balanced accuracy for predicting hypertrophy (0.84 ± 0.08), injury (0.80 ± 0.09), and proliferative lesions (0.80 ± 0.10). Though chemical and bioactivity classifiers had a similar balanced accuracy, the former were more sensitive, and the latter were more specific. CART, ENSMB, and SVM classifiers performed the best, and nuclear receptor activation and mitochondrial functions were frequently found in highly predictive classifiers of hepatotoxicity. ToxCast and ToxRefDB provide the largest and richest publicly available data sets for mining linkages between the in vitro bioactivity of environmental chemicals and their adverse histopathological outcomes. Our findings demonstrate the utility of high-throughput assays for characterizing rodent hepatotoxicants, the benefit of using hybrid representations that integrate bioactivity and chemical structure, and the need for objective evaluation of classification performance.