Evaluating Chemicals for Thyroid Disruption: Opportunities and Challenges with in Vitro Testing and Adverse Outcome Pathway Approaches

Background: Extensive clinical and experimental research documents the potential for chemical disruption of thyroid hormone (TH) signaling through multiple molecular targets. Perturbation of TH signaling can lead to abnormal brain development, cognitive impairments, and other adverse outcomes in humans and wildlife. To increase chemical safety screening efficiency and reduce vertebrate animal testing, in vitro assays that identify chemical interactions with molecular targets of the thyroid system have been developed and implemented. Objectives: We present an adverse outcome pathway (AOP) network to link data derived from in vitro assays that measure chemical interactions with thyroid molecular targets to downstream events and adverse outcomes traditionally derived from in vivo testing. We examine the role of new in vitro technologies, in the context of the AOP network, in facilitating consideration of several important regulatory and biological challenges in characterizing chemicals that exert effects through a thyroid mechanism. Discussion: There is a substantial body of knowledge describing chemical effects on molecular and physiological regulation of TH signaling and associated adverse outcomes. Until recently, few alternative nonanimal assays were available to interrogate chemical effects on TH signaling. With the development of these new tools, screening large libraries of chemicals for interactions with molecular targets of the thyroid is now possible. Measuring early chemical interactions with targets in the thyroid pathway provides a means of linking adverse outcomes, which may be influenced by many biological processes, to a thyroid mechanism. However, the use of in vitro assays beyond chemical screening is complicated by continuing limits in our knowledge of TH signaling in important life stages and tissues, such as during fetal brain development. Nonetheless, the thyroid AOP network provides an ideal tool for defining causal linkages of a chemical exerting thyroid-dependent effects and identifying research needs to quantify these effects in support of regulatory decision making. https://doi.org/10.1289/EHP5297


Introduction
Regulatory programs within the U.S. Environmental Protection Agency (EPA) are responsible for protecting public health and the environment from the potential hazards and risks of chemical exposures. To quickly and economically predict chemical hazard, and reduce laboratory animal testing, the U.S. EPA and many other government institutions and stakeholders are placing an emphasis on designing, validating, and implementing alternative approaches to whole animal toxicity testing to support chemical safety assessments and risk-based decisions (ECHA 2016;ICCVAM 2018;National Academies of Sciences, Engineering, and Medicine 2007;. These alternative approaches, more recently termed new approach methodologies, may encompass any of a broad range of in vitro technologies [e.g., robotic-based higher-throughput screening (HTS), lower-throughput formats], omic approaches (e.g., microarray, RNA-sequencing, proteomics, genome editing), and in silico modeling (e.g., molecular docking to model interactions of small molecules and proteins, computational read-across to predict the toxicity of data poor substances using data rich sources).
Coincident with the increasing availability of alternatives to traditional animal testing, government agencies have been seeking to adopt these new approach methodologies in chemical safety evaluations. The Interagency Coordinating Committee on the Validation of Alternative Methods (ICCVAM) and National Toxicology Program Center for the Evaluation of Alternative Toxicological Methods recently finalized strategies for establishing new approaches in chemical safety evaluations for U.S. federal agencies (ICCVAM 2018). The U.S. EPA has published plans to promote the use of nonanimal alternative test strategies under the amended Toxic Substances Control Act (TSCA) (U.S. EPA 2018). Likewise, the U.S. EPA's Endocrine Disruptor Screening Program (EDSP), which screens and tests chemicals for interference with estrogen, androgen, and thyroid hormone (TH) signaling, has been transitioning toward use of in vitro HTS to evaluate chemicals for endocrine activity (U.S. EPA 2015b(U.S. EPA , 2017. In vitro HTS has been implemented under the EDSP to screen chemicals for interaction with estrogen and androgen receptors (Browne et al. 2015;Judson et al. 2015Judson et al. , 2017Kleinstreuer et al. 2017;U.S. EPA 2014), as well as steroidogenesis pathways (Botteri Principato et al. 2018;Haggard et al. 2018a;Karmaus et al. 2016). Several in vitro HTS assays for identification of thyroid active chemicals are now available, or under development, and the ToxCast and Tox21 programs have implemented in vitro HTS for targets in the thyroid pathway (Collins et al. 2008;Dix et al. 2007;Judson et al. 2010;Kavlock et al. 2008; Thomas et al. 2018). In Europe, the Joint Research Commission European Union Reference Laboratory for Alternatives to Animal Testing has been developing and validating in vitro thyroid assays for test guidelines (EU Science Hub 2017), following an Organisation for Economic Co-operation and Development (OECD) scoping document on the availability of in vitro methods to investigate modulators of TH signaling (OECD 2014). Efforts to develop new methods for evaluating chemical interactions with the thyroid pathway have been informed by recent initiatives to evaluate the empirical evidence of chemical mechanisms of thyroid disruption (e.g., EU 2017; Murk et al. 2013).
To meet regulatory requirements for identifying potential endocrine disruptors, integration of a chemical's toxicity mechanisms and apical effects is needed. To date, this has been challenging for identifying chemicals that perturb TH signaling due to the limited availability of mechanistic and thyroid-relevant end points used to evaluate chemical safety. The OECD's recently updated Guidance Document 150 includes a revised conceptual framework and associated in vitro and in vivo test guidelines for evaluating endocrine disruption, including guidelines that have been updated to add thyroid end points (OECD 2018j). The European Chemicals Agency has adopted OECD Guidance Document 150 in its regulatory guidance for evaluating biocides and pesticides for endocrine disruption (Andersson et al. 2018).
Despite these recent advances, the application of new technologies to identify chemicals that exert effects through a thyroid mechanism continues to be complicated by a lack of linkages between chemical effects detected at macromolecular levels of biological organization (e.g., protein, cellular) and end points (e.g., serum TH) and adverse outcomes (e.g., developmental neurotoxicity) traditionally used in hazard assessment. Moreover, apical developmental end points collected in current regulatory test guidelines (Table 1) are not thyroid-specific and may occur through a variety of mechanisms, making it difficult to infer biological relevance, dose-and time-response relationships, and associated uncertainties. Taken together, regulatory test guidelines, when required for chemical assessments and regulatory decisions, continue to rely on thyroid measurements derived from in vivo animal testing with uncertainty about the mechanisms underlying adverse outcomes when observed.

Objectives and Overview of Challenges
To facilitate the regulatory application of data derived from these new technologies, we present an adverse outcome pathway (AOP) network (Figures 1-2) that links well-known and putative chemical targets of thyroid activity to downstream adverse outcomes. We examine current regulatory and scientific challenges in screening chemicals for thyroid-related perturbations, and propose how new in vitro technologies in the context of the thyroid AOP network can serve as a screening tool to identify chemicals that interact with targets in the thyroid pathway. Measuring chemical interactions with thyroid-related molecular targets can help elucidate whether adverse outcomes are mediated through a TH signaling mechanism.
TH signaling involves feedback interactions of the hypothalamic-pituitary-thyroid (HPT) axis, circulatory system, liver, and other tissues, collectively referred to here as the thyroid axis, that are critical to the maintenance of homeostasis and regulation of development and physiological functions (Bernal 2005;Denver 1998;Dussault and Ruel 1987;Oppenheimer and Schwartz 1997;Power et al. 2001;Silva 1995;Yen 2001;Zhang and Lazar 2000;. In contrast to other endocrine pathways, evaluation of chemical impacts on the thyroid axis requires a more involved screening strategy because chemicals perturb TH signaling through several mechanisms other than the thyroid hormone receptor (TR) (Brucker-Davis 1998;Capen and Martin 1989;Capen 1997;DeVito et al. 1999;Hurley 1998; U.S. EPA 2015a). Clinical and experimental evidence describe receptor-independent disruption of TH signaling that can result in reduced serum and tissue concentrations of TH that can lead to abnormal brain development, irreversible cognitive deficits, and other impairments, depending on the severity and timing of the TH insufficiency (as reviewed by Brucker-Davis 1998;Capen and Martin 1989;Crofton 2008;Diamanti-Kandarakis et al. 2009;Gilbert and Zoeller 2010;Gore et al. 2015;Hurley 1998;Köhrle 2008;Murk et al. 2013;Zoeller 2005).
To help link molecular targets that may mediate thyroiddependent apical effects, AOPs can be integrated in an AOP network ( Figure 1) that includes multiple molecular initiating events (MIEs) sharing one or more key events (KEs) leading to one or more adverse outcomes (Ankley et al. 2010;Browne et al. 2017;Paul Friedman et al. 2016;Villeneuve et al. 2014;Villeneuve et al. 2018;Wittwehr et al. 2017). The thyroid AOP network herein ( Figure 2) provides a tool for mapping TH signaling data collected at multiple levels of biological organization and using differing study designs. It is intended to aide in evidence integration and determining if a chemical is exerting an adverse outcome by a thyroid mechanism.

In Vitro HTS Assays Aimed at MIEs in the Thyroid Axis
Over two dozen MIEs in the thyroid axis are demonstrated or hypothesized to be potential targets of chemicals (Table 2). These MIEs have been described at length in a recent review (OECD 2014) and workshop , so they will only be briefly summarized here. As a result of these and other collaborative efforts over the years, an increasing number of in vitro HTS assays and lower-/medium-throughput assays amenable to high-throughput platforms have been developed to measure chemical interactions with MIEs that may lead to perturbed TH synthesis, delivery, metabolism, and signaling. Perhaps two of the most well-characterized MIEs in the thyroid axis involve disruption of TH production by chemical interference with iodide uptake into thyroid follicular cells by inhibition of the Na + /Isymporter (NIS) (Clewell et al. 2004;Lecat-Guillet et al. 2008;Richards and Ingbar 1959;Tietge et al. 2005Tietge et al. , 2010Wolff 1998;Wu et al. 2016) and by inhibition of thyroperoxidase (TPO) activity (Coady et al. 2010;Davidson et al. 1978;Degitz et al. 2005;Francis and Rennert 1980;Tietge et al. 2010). An in vitro radioactive iodide uptake assay coupled to a human NISexpressing HEK293T cell line has been developed and adapted for use as an in vitro HTS assay to identify chemicals that may inhibit iodide uptake by NIS (Hallinger et al. 2017;Wang et al. 2018). Another in vitro HTS assay has been developed to use the commercially available fluorescent peroxidase substrate, Amplex UltraRed to detect TPO inhibition (the "AUR-TPO" assay), and was used to screen approximately 1,000 chemicals in the ToxCast phase 1 and 2 libraries (Paul Friedman et al. 2016;Paul et al. 2014).
Extrathyroidal targets involved in chemical perturbations of TH metabolism and transport have also been the focus of in vitro HTS development. Some chemicals have been shown to inhibit the activity of iodothyronine deiodinase (DIO) enzymes in different species, tissues, and life stages (Butt et al. 2011;Capen and Martin 1989;Ferreira et al. 2002;Hood and Klaassen 2000;Morse et al. 1993;Noyes et al. 2011Noyes et al. , 2013. The three DIO isoforms, Types I, II, and III (DIO1, DIO2, and DIO3, respectively) are differentially expressed and critical to the maintenance of TH in circulation and target tissues, such as brain (Dentice et al. 2013;Gereben et al. 2008;Köhrle 1999;Salvatore et al. 1996;St Germain et al. 1994). A low-throughput colorimetric assay (Renko et al. 2012) has been adapted recently to 96-well plate Table 1. U.S. EPA and OECD test guidelines with required (X) and optional (opt) in vivo thyroid end points and potentially thyroid-responsive adverse outcomes; laid out according to the adverse outcome pathway (AOP) diagram in Figure 1.
Test format using an adenovirus that expresses human DIO1, and was optimized to detect DIO1 inhibition by chemicals in the ToxCast phase 1 library (Hornung et al. 2018). This assay is currently being used to design in vitro HTS assays to evaluate chemical inhibition of DIO2 and DIO3 (Olker et al. 2019). Chemicals also may act as ligands that can bind TH distributor proteins, notably transthyretin (TTR), and to a lesser extent thyroid binding globulin (TBG), in serum and cerebrospinal fluid (CSF) to displace native thyroxine (T4) (Brouwer et al. 1986;Cheek et al. 1999;Ishihara et al. 2003;Meerts et al. 2000;Ren and Guo 2012;Weiss et al. 2009). Although the overall in vivo relevance of chemical interference with serum TH distributors remains unclear, a surface plasmon resonance-based biosensor assay for TTR and TBG is available that provides medium-to high-throughput testing capabilities with commercially available technologies (Marchesini et al. 2006(Marchesini et al. , 2008. Another mechanism by which chemicals decrease circulating concentrations of TH is by activation of hepatic xenobiotic nuclear receptors (NRs) leading to inductions of phase I, II, and III metabolic enzymes and transporters in the liver and other tissues. Enhanced phase II TH glucuronidation and sulfation catalyzed by uridine diphosphate glucuronosyltransferases (UDPGTs) and sulfotransferases (SULTs), respectively, can increase TH catabolism and reduce serum TH by accelerating clearance (Barter and Klaassen 1994;Hood et al. 2003;Yu et al. 2009;Zhou et al. 2002). In vitro HTS assays are available in ToxCast to assess chemical binding and activation of specific xenobiotic NRs [e.g., constitutive androstane receptor (CAR); pregnane X receptor (PXR); aryl hydrocarbon receptor (AhR)]. Once bound, activation of these NRs may upregulate expression of phase I (CYP450s) and phase II [e.g., UDP glucuronosyltransferase family 1 member A1 and member A6 (UGT1A1 and UGT1A6, respectively), Sulfotransferase family 2A member 1 (SULT2A1)] genes encoding isoenzymes involved in T4 glucuronidation and sulfation. In vitro HTS assays measuring induction and/or inhibition of TH conjugating enzymes, as well as the transporters mediating cellular transport of TH, are under development.
With regard to receptor-ligand interactions, screening of the Tox21 chemical library of 10,000 chemicals using transactivation assays aimed at TRα and TRβ indicate that binding is restricted to a limited number of chemicals (Freitas et al. 2011(Freitas et al. , 2014Houck et al. 2018; U.S. EPA 2015a). Likewise, chemical interactions with the thyrotropin releasing hormone receptor (TRHR) and thyroid stimulating hormone receptor (TSHR) do not appear to be prominent sites for chemical binding (Gershengorn and Neumann 2012;Murk et al. 2013); however, this is an area with limited research and analysis of the ToxCast/Tox21 screening data is underway Shobair et al. 2019).

Thyroid AOP Network as Framework for Chemical Screening and Assessment
The thyroid AOP network in Figure 2 illustrates how in vitro HTS assays (under development or currently available) identified in Table 2 can be mapped to this network of KEs that chart a progressive path to adverse outcomes. The thyroid AOP network serves as the foundation to organize and evaluate thyroid data, identify data gaps, and examine the evidence for causality between KEs in AOPs. These KE relationships connecting MIEs to KEs and adverse outcomes in Figure 2 may be considered hypothesized, correlative, or causal depending on the strength of the evidence (e.g., Coady et al. 2017;Crofton and Zoeller 2005;Degitz et al. 2005;Hassan et al. 2017;Miller et al. 2009;Murk et al. 2013;Perkins et al. 2013;Zoeller and Crofton 2005).
Another important application of the thyroid AOP network is to clarify research needs for decision making. In vitro HTS assays currently available to measure MIEs are shown in Figure 2 with solid borders in the left-hand column (and also shaded green), and end points collected as part of U.S. EPA and OECD guidelines (Table 1) are shown with thick red borders. Several MIEs in Figure 2 and further summarized in Table 2 do not yet have in vitro HTS assays even though in some instances lower-throughput in vitro assays are available (e.g., for UDPGT/SULT activity). In addition, results from studies published in peer-reviewed literature may be another source of data to further populate the AOP network. The peer reviewed literature may be particularly helpful in identifying and characterizing KEs downstream of serum TH and proximal to adverse outcomes as current regulatory test guidelines (Table 1) provide limited coverage.
Advancements of methods for systematic review and evidence integration provide approaches to identify and evaluate peerreviewed studies for relevance, performance, and reliability to support their inclusion in chemical assessment (Hoffmann et al. 2017;Rooney et al. 2014; National Academies of Sciences, Engineering, and Medicine 2018). An increasing number of omic studies also Figure 1. An adverse outcome pathway (AOP) begins with a molecular initiating event (MIE) and terminates in an adverse outcome that is linked by a series of intermediate key events (KEs) at increasing levels of biological organization. Adverse outcomes at the organism level are used in human health risk assessment and typically with plausible linkages to the population level for ecological risk assessment. A simplified example of an AOP network is presented whereby three MIEs shown in the first column (shaded green) elicit specific cellular responses in the next three columns (KE1, KE2, KE3; shaded orange) that converge in shared organ pathology (KE4) and mediate downstream organ system alterations (KE5) to produce divergent adverse outcomes at the organism (AO1, AO2) and population (AO3) levels as shown in the last two columns (shaded red). describe chemical effects on genes (transcriptomics), proteins (proteomics), and low-molecular-weight biomolecules (metabolomics) in the thyroid gland and TH-responsive tissues, including several transcriptomic studies (Boucher et al. 2014;Crump et al. 2002;Huang et al. 2011;Haggard et al. 2018b;Ohara et al. 2018;Porreca et al. 2016;Royland et al. 2008). For example, chromatin immunoprecipitation with microarray (ChIP-on-chip) and DNA sequencing (ChIP-seq) have been used in genome-wide analyses to identify putative TR target genes in human cancer cell lines (Chung et al. 2016), mouse liver (Grøntved et al. 2015;Ramadoss et al. 2014), mouse brain (Compe et al. 2007;Dong et al. 2009), and frog intestine (Fu et al. 2017). There are now a handful of studies reporting chemical effects on thyroid-related proteomics (e.g. metabolomics (e.g., SSY Huang et al. 2016;Houten et al. 2016, Johnson et al. 2012. Cell-based assays and transgenic animal models have also been deployed to elucidate molecular targets, toxicity pathways, and chemical structure-activity relationships of thyroid disruption (Gentilcore et al. 2013;Jarque et al. 2018;Ji et al. 2012;Opitz et al. 2012;Rosenberg et al. 2017). To this end, recent advances in genome editing tools with toxicological applications, notably clustered, regularly interspaced, short palindromic repeats (CRISPR)-Cas9 technologies have been used to better understand TH signaling pathways (Kyono et al. 2016;Markossian et al. 2018;Sakane et al. 2018;Trubiroha et al. 2018;Yang et al. 2018). However, despite these advances, inclusion of in vitro genotyping studies in chemical assessment is still rare, and current efforts seek to standardize methods and link genetic outcomes to phenotypic  In the left-hand column, MIE boxes with solid borders (shaded green) represent current MIEs with in vitro high-throughput screening (HTS) assays that have demonstrated reliability and are available for use in thyroid activity screens, whereas those with dashed borders represent putative MIEs in the thyroid axis currently without in vitro HTS capabilities. In the key events (KEs) column, the box with the striped background (shaded yellow) depicts changes in serum TH as a KE node that represents a biomarker of thyroid disruption, whereas the trapezoids (shaded blue) represent additional potential KE nodes with limited data. Uppercase nomenclature denoting human protein is shown although present in differing species. Asterisks represent KEs being treated as MIEs. References and AOPs supporting the thyroid AOP network are identified in Table 2. AhR, aryl hydrocarbon receptor; BDNF, brain-derived neurotrophic factor; CAR, constitutive androstane receptor; DIO, iodothyronine deiodinase; DIO1, type 1 deiodinase; DIO2, type 2 deiodinase; DIO3, type 3 deiodinase; DUOX, dual oxidase; IYD, iodotyrosine deiodinase; LDL, low-density lipoprotein; MDR, multidrug resistance protein; MCT, monocarboxylate transporter; NIS, sodiumiodide symporter; OATP, organic anion transporter polypeptide; OECD, Organisation for Economic Co-operation and Development; PPAR, peroxisome proliferator-activated receptor; PXR, pregnane X receptor; rT3, reverse T3 (3,3 0 ,5 0 -triiodothyronine); RXR, retinoid X receptor; SULT, sulfotransferase; T3, 3,3 0 ,5triiodothyronine; T4, thyroxine; TBG, thyroid binding globulin; TH, thyroid hormone; TPO, thyroperoxidase; TR, thyroid hormone receptor; TRHR, thyrotropin releasing hormone receptor; TSHR, thyroid stimulating hormone receptor; TTR, transthyretin; UDPGT, uridine diphosphate glucuronosyltransferase. In vitro studies: Davidson et al. 1978 Reviews: Dellarco et al. 2006;Hurley 1998 Iodotyrosine deiodinase (IYD) Scavenges/recycles iodide in the thyroid by catalyzing deiodination to T1 and T2. Limited evidence of chemical inhibition. The toxicological mechanism column highlights the role of MIEs in thyroid hormone (TH) signaling along with chemicals shown to interact with them. b In vitro HTS assay development denoted as "promising" indicates MIEs in the thyroid axis for which there is interest and/or activity in developing in vitro HTS approaches, typically with supportive in vivo and slow-or medium-throughput in vitro toxicity studies indicating chemical interactions. In vitro HTS readiness denoted as "early" indicates putative MIEs but with limited toxicity evidence and with little current activity to develop high-throughput alternatives.
c MIEs in the thyroid axis with evidence of linkages to adverse outcomes. Additional information on individual AOPs in development and completed can be found at https://aopwiki.org/. Potential adverse outcomes c

References
Type 2 (DIO2); DIO Type 3 (DIO3) been shown to induce thyroid tumors in rats, no studies to date have shown chemicals that exert effects on DIO expression and/or activity to directly manifest in adverse outcomes. Morse et al. 1993;Noyes et al. 2011Noyes et al. , 2013Szabo et al. 2009 In  Neumann et al. 2009;Santini et al. 2003;Titus et al. 2008 In silico studies: Paul Friedman et al. 2017;Shobair et al. 2019 Reviews: Gershengorn and Neumann 2012

Chemicals: Unknown
TR binding and transactivation (TRα, TRβ) Transcription factors that have ligand (T3)-dependent and -independent activity. In humans, THRA genes encode TRα1, TRα2, and TRα3 and truncated isoforms. THRB genes encode TRβ1, TRβ2, and TRβ3 (possibly rat only) and truncated isoforms. Only TRα1, TRβ1, TRβ2, and TRβ3 can bind T3 and TREs; TRβ1 and TRβ2 regulate TRH in the hypothalamus. Some chemicals bind TRs as antagonists and/ or modify transcription; however, screens of chemical libraries suggest binding is restricted.

Existing:
ToxCast/Tox21: Freitas et al. 2011Freitas et al. , 2014 Not yet characterized In vivo studies: Dupré et al. 2004 In vitro studies: Cheek et al. 1999;Clerget-Froidevaux et al. 2004;Freitas et al. 2011Freitas et al. , 2014Gauger et al. 2007;Hofmann et al. 2009;Huang et al. 2011;Kitamura et al. 2002Kitamura et al. , 2005Kojima et al. 2009;Moriyama et al. 2002;Schriks et al. 2006;Sun et al. 2009;You et al. 2006 In silico studies: Knudsen et al. 2011;Politi et al. 2014;Romanov et al. 2008;Sipes et al. 2013Reviews: Zoeller 2005 Chemicals: TBBPA, TCBPA, BPA, OH-PCBs, OH-BDEs, triclosan anchors representing adverse outcomes (Buesen et al. 2017;Bourdon-Lacombe et al. 2015;Brazma et al. 2001;Dean et al. 2017;Thomas et al. 2013aThomas et al. , 2013bVilleneuve et al. 2012;Wang et al. 2016). Assembling and interpreting the scientific evidence measured in independent in vitro, in silico, and in vivo research efforts presents an enormous challenge for chemical assessment. The thyroid AOP network provides architecture to begin mapping and integrating diverse data to characterize chemical interactions with the thyroid axis. New in vitro technologies incorporated into pathway-based approaches are anticipated to help reduce uncertainties and resolve some of the major challenges that are examined below and presented by these evaluations.  In contrast, elevated TSH (shown with a dashed border) has not been shown as a causative KE in chemical inhibition of TPO leading to cognitive impairments in mammals (B) or chemical activation of the hepatic xenobiotic NR response cascade leading to mammalian hearing deficits (C). References supporting these AOPs can be found in Table 2. $, no change from reference controls; AhR, aryl hydrocarbon receptor; CAR, constitutive androstane receptor; KE, key event; MIE, molecular initiating event; NR, nuclear receptor; PXR, pregnane X receptor; T3, 3,3 0 ,5-triiodothyronine; T4, thyroxine; TPO, thyroperoxidase; TH, thyroid hormone; TR, thyroid hormone receptor; TRH, thyrotropin releasing hormone; TRHR, thyrotropin releasing hormone receptor; TSH, thyroid stimulating hormone; TSHR, thyroid stimulating hormone receptor; UDPGT, uridine diphosphate glucuronosyltransferase.

Regulatory Challenges in Evaluating and Linking Thyroid MIEs to Adverse Outcomes
Currently, no U.S. EPA or OECD in vitro regulatory guidelines test chemical interactions with MIEs in the thyroid axis (Table 1). For most regulatory requirements, hazards are identified for individual chemicals and there are little to no data describing the combined effects of individual chemicals and chemical mixtures that interact with multiple MIEs in the thyroid axis. In addition, current guidelines can detect some thyroid-related disturbances (e.g., changes in serum TH) that have been used to derive hazard values for regulatory purposes, but it is not possible to attribute an adverse outcome to a mechanism. Neither have intermediate parameters (KE and KE relationships) linking an adverse outcome to an upstream MIE been empirically quantified in most cases. By depicting points where several MIEs converge at shared KEs, the AOP network helps to identify nodes around which assays can be developed to link MIEs to adverse outcomes and clarify data that would be most relevant for regulatory applications. For example, uncertainties continue as to dose-response relationships in chemical-induced changes in serum TH, particularly milder TH perturbations that result in adverse outcomes (Crofton 2004). For example, AOPs ( Figure 2; Table 2) describe chemical inhibition of NIS and TPO leading to reduced TH production and decreased serum T4 that culminate in delayed/ arrested amphibian development (AOP 176, AOP 175, respectively) and impaired mammalian neurodevelopment (AOP 134, AOP 42, respectively). However, a quantitative understanding of KE relationships in the TPO-and NIS-related AOPs is limiting, although current efforts are beginning to quantify these interactions (Gilbert and Sui 2008;Hassan et al. 2017;O'Shaughnessy et al. 2018b). Decision makers can use the biological plausibility embedded in the thyroid AOP network to clarify linkages for characterization and identify gaps in available knowledge. From this information, KE relationships can be proposed for quantification to support predictive modeling and hazard assessment, as well as reduce data uncertainties.

Differential Species and Life Stage Sensitivities; Interpreting Biomarkers
Future work to expand and refine the thyroid AOP network can help to identify physiological similarities and differences between KEs and KE relationships spanning vertebrate classes. The characterization of shared toxicity mechanisms leading to divergent adverse outcomes across species can be useful in the translation of research between nonmammalian and mammalian species including humans. For example, chemical inhibition of thyroid gland TPO can reduce serum and tissue TH that alter tissuespecific genomic signaling and elicit species-and life stagespecific adverse outcomes [e.g., abnormal brain development in mammals ; impaired metamorphosis in amphibians (Coady et al. 2010;Hornung et al. 2015;Tietge et al. 2010); reduced anterior swim bladder inflation in teleost fish (Nelson et al. 2016;Stinckens et al. 2016)]. Identification of shared KEs as integrative nodes facilitates characterization of diagnostic biomarkers across species that can be represented in predictive modeling and evaluated in organismal assays to further support mechanistic linkages between KEs and adverse outcomes (Crofton 2008;Miller et al. 2009;Perkins et al. 2013).
Several MIEs in the thyroid AOP network may induce a cascade of KEs proceeding through decreased serum T4 and/or triiodothyronine (T3) (Figure 2; box with striped background and shaded yellow). Changes in serum TH are a useful biomarker of thyroid perturbation and are usually correlated to changes in tissue TH (as reviewed by Oppenheimer and Schwartz 1997;Porterfield and Hendrich 1993). However, reliance on serum TH may oversimplify subsequent linkages to downstream KEs that drive adverse outcomes. Serum TH levels are not always aligned with tissue levels as tissue TH is regulated by DIO enzymes, cellular transporters, and regional TR expression. The disconnect between altered TH signaling in the fetal brain, neurodevelopmental consequences, and serum TH has been demonstrated using several gene knock-out (KO) mouse models with deficiencies in genes encoding DIO enzymes, TRs, monocarboxylate transporter (MCT)8, organic anion transporter polypeptide (OATP)1c1, and other TH transporters (as reviewed by Bárez-López et al. 2018;Bernal 2018;Hernandez et al. 2010;Richard and Flamant 2018). For example, considerable evidence describes the contribution of DIO enzymes, particularly DIO2 and DIO3, in the maintenance of T3 in THresponsive regions of the developing brain [i.e., DIO2 predominantly in astrocytes; DIO3 predominantly in neurons; as reviewed by Bianco et al. (2002), Guadaño-Ferraz et al. (1997), Morreale de Escobar et al. (2008), Patel et al. (2011), andZoeller (2010)]. However, studies show that Dio2 and Dio1/Dio2 KO mouse models with deficient tissue capacity to produce T3 do not experience substantial deficits in motor, learning, or memory functioning, although other deficits (e.g., impaired cochlear and visual development) do occur (Galton et al. 2007(Galton et al. , 2009(Galton et al. , 2014Obregón et al. 1991;Schneider et al. 2001;Schneider et al. 2006). These Dio1/ Dio2 KO strains exhibit regionally specific reductions in brain T3 but normal serum T3, as well as elevated serum and tissue T4. In contrast, the Dio3 KO fetus and newborn pup exhibit an initial hyperthyroidism followed by hypothyroidism, impaired growth and fertility, and reduced survival. Elevations in serum and brain T3 are accompanied by an initial acceleration and subsequent delay of T3-inducible gene expression in the brain (Hernandez et al. 2010). Subsequent experiments have shown that Dio3 KO mice with augmented brain TH exhibit hyperactivity and reduced anxiety-like behaviors despite systemic hypothyroidism as indicated by reduced serum T3 (Stohn et al. 2016).
As these examples illustrate, the lack of change in serum TH is not a guarantee that TH effects in tissues are not occurring, and conversely changes in serum TH are not always predictive of adverse outcomes. Possible explanations for the lack of concordance between serum TH and adverse outcomes may relate to the not yet fully characterized age-and regional-specific responses of the brain to compensate to systemic TH insufficiency. Nonetheless, fetal and early postnatal susceptibilities to thyroid toxicants are heightened because their thyroid feedback systems are absent or incompletely formed and they have low TH reserves that are critical to neurological development (as reviewed by Morreale de Escobar et al. 2004a;Skeaff 2011;Williams 2008;and Zoeller and Rovet 2004). The thyroid AOP network provides a tool to map linkages and identify possible KEs for further characterization and quantification (e.g., trapezoid boxes also shaded blue in Figure 2) to aide in predicting adversity. For purposes of decision making, changes in serum TH profiles are not proscriptive to either mechanisms or outcomes, but they nevertheless do provide clear indication of perturbed thyroid homeostasis with the potential to adversely affect development.
The thyroid AOP network also can identify KEs that may not fall on the causative pathway leading to an adverse outcome, making these KEs potentially less relevant to decision making. One such example is the stimulation of the thyroid-pituitary feedback response (i.e., increased TSH). Although TSH increases often accompany exposure to a number of thyroid disrupting chemicals, TSH is not a causative KE in AOPs involving TPO inhibition that result in impaired hippocampal development and cognitive deficits  and up-regulated TH catabolism leading to ototoxicity   (Figure 3). For example, rat offspring developmentally exposed to some polychlorinated biphenyls (PCBs) and dioxins experience mainly hypothyroxinemia (i.e., reduced serum T4 with TSH in reference control ranges) that culminate in impaired cochlear development and hearing loss (Crofton et al. 2000;Crofton 2004;Goldey et al. 1995;Morse et al. 1993;Sher et al. 1998). Elevated TSH in rodents leads to thyroid hypertrophy and potential thyroid cancer, an adverse outcome that has limited relevance to human thyroid cancer due to species differences in sensitivity (Capen and Martin 1989;EU 2017;Hurley 1998;McClain et al. 1988), although this too is an area of renewed interest (EU 2017). Thus, elevated TSH may not be a reliable KE biomarker for thyroid cancer in humans, but increased TSH in rodents does indicate a chemical capable of perturbing TH signaling in other species, including humans, and is involved in other thyroid-responsive outcomes, including neurodevelopment.
Understanding pathway conservation is critical for cross-species extrapolation and requires understanding the conservation of protein targets (e.g., MIEs) and resulting sensitivity to chemicals. Research efforts in this area range from in vivo exposures across phylogenetically disparate species, followed by toxicogenomic meta-analyses (Garcia-Reyero et al. 2011), to bioinformatics comparing amino acid sequences/functional domain similarity across species to identify putative functional orthologues . For purposes of chemical screening and in vitro HTS assay development, a basic demonstration of cross-species concordance of chemical activity is sufficient to provide confidence that the representative in vitro system is reliable for generating screening-level data. For example, strong cross-species concordance in the activity of porcine and rat TPO has been shown despite only 75% amino acid similarity in the conserved peroxidase domain . Results from numerous in vivo exposures to select TPO-active chemicals in fish (Doerge et al. 1998;Nelson et al. 2016;Stinckens et al. 2016), amphibians (Coady et al. 2010;Hornung et al. 2015;Tietge et al. 2010Tietge et al. , 2013, birds (Grommen et al. 2011;Rosebrough et al. 2006), and mammals (Davidson et al. 1978;Divi and Doerge 1996;Francis and Rennert 1980) corroborate the conserved mechanistic linkages leading to species-specific adverse outcomes. Such information can define circumstances under which screens for thyroid activity can benefit from combining lines of evidence across species with shared KEs, as opposed to situations when species-specific data support being separated for regulatory use.

Human Relevance of Hepatic-and Serum Distributor-Driven TH Effects in Rodents
The observation of thyroid-related end point changes in rodents and their relevance to predicting adverse outcomes in humans continues to present research questions. One such data gap concerns interpreting the human applicability of chemical-induced TH reductions in rodents through activation of xenobiotic NRs, notably PXR and CAR, and up-regulation of hepatic TH catabolism (Figures 2-3; Table 2) (EU 2017;ECHA 2016). Although functionally conserved in many cases (e.g., CYP450s), hepatic drug metabolizing enzymes have differing isoform compositions, substrate selectivity, and catalytic activities depending on species, life stage, and genetics that may affect chemical and TH biotransformation (Benedetti et al. 2005;Curran and DeGroot 1991;Kondo et al. 2017;Martignoni et al. 2006;Nebert and Gonzalez 1987).
Increased hepatic catabolism of TH, along with reductions in serum TH and elevated TSH, are well described in rodents (rat, mouse, hamster) exposed to hepatic enzyme-inducing chemicals such as phenobarbital, some PCBs (Aroclor 1254), pregnenolone-16α-carbonitrile, and 3-methylcholanthrene ( Barter and Klaassen 1994;Haines et al. 2018;Hood et al. 2003;Kato et al. 2010;Klaassen and Hood 2001;Vansell and Klaassen 2002;Wong et al. 2005). Studies in human subjects also demonstrate the responsiveness of TH catabolic pathways to hepatic enzyme inducers (Benedetti et al. 2005;Christensen et al. 1989;Curran and DeGroot 1991;Ohnhaus et al. 1981;Rootwelt et al. 1978;Yeo et al. 1978;Zhang et al. 2016). In contrast with rodents, human serum T4 decrements reported in these studies are varied and typically not accompanied by changes in TSH, suggesting the possibility of milder responses in humans compared with rodents, although this may not be the case in sensitive subpopulations. For example, studies in healthy human test subjects receiving therapeutic doses of the human CAR activator phenobarbital and other antiepileptic drugs show unchanged or decreased serum TH, with generally no TSH alterations (as reviewed by Benedetti et al. 2005;Curran and DeGroot 1991). Similarly, healthy human subjects receiving therapeutic doses of the antibiotic and human PXR agonist rifampicin exhibit reduced serum T4, as well as increased T4 and rT3 clearance, with no change in TSH (Christensen et al. 1989;Ohnhaus et al. 1981). However, a meta-analysis of published clinical studies that examined TH end points in epileptic patients treated with phenobarbital and other antiepileptic drugs detected an overall decrease in serum T4 (no change in T3) and increase in TSH (Zhang et al. 2016), suggesting genetics and health status may influence sensitivity. The antibacterial triclosan, known to decrease serum TH in rats, has also been shown to activate human PXR, but not rat or mouse PXR, and to behave as an inverse agonist to human CAR1 and weak agonist of human CAR3 and rat/mouse CAR ). Thus, current evidence supports that chemicals may induce hepatic TH catabolism in both humans and rodents, albeit by possibly differing metabolic pathways and potencies (an unknown at this time). Whether responses in humans extend to other chemicals as a general phenomenon is unclear, and differing exposure durations and responses between rodents and humans may be relevant to extrapolations from one species to another. Useful follow-up analyses of chemicals suspected of NR activation could involve evaluation of species selectivity of the NR response to the chemical itself.
It has also been posited that TH metabolic clearance responses in rats are less relevant to humans because TBG, the major serum TH distributor protein in humans, is less prominent in the adult rat, although it does play an important role in thyroid homeostasis in newborn and neonatal rat pups (Savu et al. 1991;Young et al. 1988). In adult rodents, amphibians, birds, and fish, TTR appears to be the major serum TH distributor protein (Dickson et al. 1985;Harms et al. 1991;Power et al. 2000). TTR has a lower binding affinity for T4 than TBG, and this is thought to be responsible for the shorter T4 half-life in rats (∼12-24 h) than in humans (∼5 d) (Lewandowski et al. 2004). Unlike TBG, TTR is important in the transport of T4 across the blood-placental (Landers et al. 2009;Mortimer et al. 2012), blood-brain (Chanoine et al. 1992;Schreiber et al. 1990), and CSF-brain (Richardson et al. 2018) barriers. Thus, a role for human TTR in TH storage and transport appears critical during fetal development. Some researchers have also suggested that although TBG binds most TH in human circulation, TH dissociation rates from TTR and capillary transit times make TTR a significant distributor protein of TH in humans as well as in rodents (Alshehri et al. 2015;Mendel 1989). Taken together, current evidence supports the activation of putative MIEs and KEs in rats, encompassing xenobiotic NR activation, inductions of UDPGTs, SULTs, and TH distributor proteins are relevant pathways for consideration. In this regard, the rat serves as a conservative model for predicting human health effects. The thyroid AOP network allows for the integration of new in vitro mechanistic and interspecies scaling information to help relate rodent toxicity data to human-equivalent exposures and further populate the TH catabolic pathway for use in human health assessment.

Building Predictive Models and Toxicokinetics
Because of the diversity of known and putative molecular targets in the thyroid axis, using AOPs to develop predictive models of adverse outcomes from in vitro data will be challenging and require additional bioassay data for KEs in the network, particularly downstream of serum TH. Although there is qualitative evidence linking the MIEs listed in Table 2 to altered serum TH, data to establish quantitative KE relationships helpful for regulatory application are lacking. Furthermore, the large diversity of genes regulated directly and indirectly by T3, the nongenomic effects of T4, and our limited knowledge of the functioning of TH during fetal development present a challenge for predicting chemical-induced adverse outcomes from in vitro data (as reviewed by Bernal 2018; Richard and Flamant 2018).
In humans, detrimental effects of mild TH insufficiency on fetal cognitive development are recognized as an important public health concern, and some links to chemical exposures are described (Finken et al. 2013;Ghassabian et al. 2014;Haddow et al. 1999;Korevaar et al. 2016;Moleti et al. 2011;Morreale de Escobar et al. 2000;Steinmaus et al. 2016;Vermiglio et al. 2004). The observations in humans are supported by rodent studies where neurodevelopmental impairments result from moderate and transient reductions in serum TH induced by chemical exposures (Ausó et al. 2004;Boyes et al. 2018;Gilbert 2011;Gilbert et al. 2013Gilbert et al. , 2016Gilbert et al. , 2017Morreale de Escobar et al. 2004b). As quantitative data become available for these types of sensitive developmental responses-likely from the use of bioassays of greater biological complexity that integrate multiple MIEs and KEs-additional modeling approaches to further characterize KE relationships could be implemented to predict thyroid-related adverse outcomes and characterize accompanying uncertainties. As a first step, models that integrate in vitro assay results for a single KE (when multiple assays exist) and models that integrate in vitro assay results for more than one KE in a target tissue (e.g., thyroid gland; TPO, NIS, DIO) could be used to prioritize chemicals for further assessment and targeted in vivo testing. Development of a computational systems biology model of thyroid function could help simulate the potential interactions of chemicals with activity at various MIEs. Such an effort will require an investment of research to parameterize and evaluate uncertainties.
Another challenge is that chemical toxicokinetic considerations, and an understanding of potential differences among life stages and species, are not well accounted for by current in vitro test systems. These limitations can lead to both false positive and false negative calls in screening efforts. Developing in vitro predictions of thyroid perturbations is further complicated by the fact that most chemically induced effects on downstream KEs and adverse outcomes are indirect, secondary consequences to a chemically induced change in hormone regulation. We think predictions of in vivo adverse outcomes from in vitro HTS data and other new technologies will be greatly facilitated by a) incorporation of toxicokinetic tools to estimate metabolic bioactivation and better characterize potency and selectivity of chemicals for target tissues, and b) improved mechanistic and quantitative understanding of TH action on development of target organs, especially the brain.

Conclusions
Advancements in new in vitro technologies offer opportunities to screen chemicals for interactions with MIEs in the thyroid pathway and to anchor adverse outcomes to a thyroid mechanism. A substantial body of knowledge exists concerning the molecular and physiological regulation of TH signaling and responses of these systems to chemical exposures (including pharmaceuticals). The thyroid AOP network described in the present work is based on decades of research and multiple consortia to evaluate relevant KEs for chemical-induced thyroid disruption. At present, the use of in vitro data beyond screening and prioritization for thyroid bioactivity is challenged by the complexity of potential THrelated adverse outcomes and the limited knowledge of mechanistic processes controlling such responses. Despite these knowledge gaps, we propose that the thyroid AOP network provides the necessary biological structure to organize diverse data from differing test methods, thereby serving as a useful data integration tool to assist in hazard screening for large numbers of previously untested chemicals. As more data become available, the AOP network can be further populated to help bridge knowledge gaps and identify key nodes for assay development.