The U.S. EPA’s ToxCast program (Kavlock et al., 2012) and cross-agency Tox21 program (Tice et al., 2013) are building large collections of in vitro data on diverse sets of chemicals to which humans are potentially exposed, including pesticides, food, cosmetics and personal care ingredients, pharmaceuticals, and industrial chemicals. Chemicals are being tested for bioactivity at various levels of biological organisation in a broad battery of in vitro assays that include cell-free systems, cell lines and primary cells from multiple tissue types, complex culture systems, embryonic stem cells and zebrafish embryos. The ToxCast database can be found at http://epa.gov/ncct/toxcast/data.html (release date December 2014) and explored by chemical or assay using the Interactive Chemical Safety for Sustainability (iCSS) dashboard (http://actor.epa.gov/dashboard/).
The utility of ToxCast data in AOPs for developmental toxicity was demonstrated by Sipes et al. (2011), who built a predictive model in which the in vitro high-throughput screening data (ToxCastDB) was anchored to in vivo adverse outcomes from prenatal developmental toxicity studies (ToxRefDB). This early model utilised the first phase (Phase-I) of ToxCast, which consisted of 309 chemicals, mostly pesticide compounds, and a range of over 600 high-throughput screening assays. In the analysis, univariate associations (one assay to one endpoint) were used to identify significant in vitro to in vivo correlations. Multivariate predictive models (multiple assays to multiple endpoints) were then built from these identified assays using linear discriminant analysis with five-fold cross validation. The rat developmental toxicity multivariate model had a predictive accuracy of 71% (sensitivity 0.72, specificity 0.70; P-value = 7.5E-11). Among the positive predictors composing this model, the RAR assay set was the strongest weighted variable (weight factor 0.58) followed by G protein-coupled receptors (weight factor 0.55), TGF-beta (weight factor 0.38), microtubule organisation (weight factor 0.30), and other lesser weighted features (Sipes et al., 2011).
Since the Sipes et al. (2011) study, ToxCastDB has expanded to include in vitro results for 1,858 chemicals and up to 821 assay features. The latter derives from 541 unique high-throughput screening assays that can be mapped to 293 molecular targets and high-throughput screening assays for diverse cellular behaviours and responses, including 37 different assays for cytotoxicity (Judson et al., 2016). Several recent analyses of the ToxCastDB (in vitro) and ToxRefDB (in vivo) data identified the retinoid pathway as a major component in models for male reproductive developmental defects (Leung et al., 2015), cleft palate (Baker et al. in preparation), and digital defects (Ahir et al. 2014, and in preparation). Since RA signalling mediates proper growth and differentiation of the embryo, a potential application for ToxCast is to identify possible targets that could, in the context of AOPs, define MIEs for critical alterations to RA homeostasis or signalling pathways. Across the 293 molecular targets, gene ontology (GO) annotations produced 18,367 records of which at least 52 could be mapped to a molecular target in the retinoid system. This includes reporter assays for transactivation of retinoid receptors (RARs and RXRs), and cis-activation of the retinoic acid response element (RARE) by RAR/RXR heterodimers. A detailed analysis of chemical-bioactivity profiles for in vitro targets in the retinoid signalling system is presently underway; however, for the purposes of this illustration, we simply mined for coherent linkages to MIEs affecting the aforementioned assay targets. All-trans retinoic acid (ATRA) was a reference compound used to benchmark the AC50 (concentration at 50% of maximum activity) for each particular assay.
A total of 879 ToxCast AC50s were mapped to a molecular target in the retinoid system (Baker et al., 2016). In total, 97 of 1,858 chemicals (5.2%) hit one or more assays in the retinoid system at an AC50 below 2 uM. With regards to retinoid metabolism (KEGG pathway hsa00830: Retinol metabolism), the ToxCast dataset presently lacks information on retinal dehydrogenase (EC: 22.214.171.124; RALDH), the enzyme that generates RA from retinol, and on cytochrome-P450 family 26 (EC: 1.14.-.-; CYP26), the enzyme specific to its breakdown. However, the dataset does have results on the biochemical activity of cytochrome-P450 family 1, subfamily A, polypeptide 1 (EC:126.96.36.199; CYP1A1), another enzyme capable of ATRA breakdown. ATRA competed with the substrate of the CYP1A1 assay to inhibit its biochemical activity with an AC50 of 1.32 uM, whereas retinol was inactive. Flusilazole, an antifungal known to disrupt RA homeostasis and invoke dysmorphogenesis (Tonk et al., 2015), registered an AC50 of 3.7 uM on CYP1A1 activity, and, in all, 11 ToxCast chemicals, mostly pesticides, inhibited CYP1A1 activity at AC50s below ATRA. This supports other evidence that disruption of RA homeostasis is a possible MIE for developmental AOPs and that some environmental chemicals may disrupt normal development through this mechanism.
ToxCast has reporter gene assays for three distinct RAR subtypes (RARα, RARβ, RARγ) and three distinct RXR subtypes (RXRα, RXRβ, RXRγ). The main assay platform utilises a HepG2 cell line engineered for transactivation of reporter genes and fold-induction in response to chemical exposure. ATRA was the most potent of all chemicals tested in the RARα and RXRα transactivation assays (subnanomolar AC50 values of 0.429 nM and 0.309 nM, respectively). Retinol had weaker AC50 values of 69 nM (RARα) and 1.54 µM (RXRα). RXR/RAR heterodimers bind to RAREs composed of tandem 5’-AGGTCA-3’ sites known as DR1-DR5; ATRA activated the DR5 cis-reporter assay with an AC50 value of 6.26 nM whereas retinol had a moderate AC50 of 147 nM. Across the entire ToxCast inventory the numbers of chemicals registering an AC50 < 2 µM were: 9 (RARα), 4 (RARβ), 6 (RARγ), 9 (RXRα), 23 (RXRβ), 0 (RXRγ), and 51 (DR5).
Some classes of persistent organic pollutants (POPs) preferentially activated the RARs at AC50s below 2 uM (e.g., organochlorines). In contrast, at least two classes of environmental chemicals preferentially activated RXRs with AC50s below 2 µM (e.g., tert-butyl compounds) or 0.2 uM (e.g., organotins). Those compounds displayed similar activity on the DR5 assay as well as distinct retinoid receptors. A number of pesticides that disrupt mitochondrial respiration displayed activity on DR5 responses with AC50s below 2 uM (e.g., strobins, rotenone). Thus, in vitro profiling of the retinoid signalling system in ToxCast identified approximately 5% chemicals with a potential for direct disruption of RA signalling through transactivation of RAR or RXR systems at submicromolar concentrations. Given the potential for these receptors to heterodimerise with different nuclear receptor subtypes (e.g. RARα with RXRα; RXRα with PPARγ, LXRβ, VDR, TRβ, or FXR), the analysis of ToxCast data allows a provisional catalogue of MIEs to be built that mechanistically invoke AOPs associated with RA signalling and homeostasis pathways. An ontology for developmental toxicity is necessary to put this complexity into a computable and integrated form.
Chemical risk assessment is at a crossroads, moving from classical animal studies looking for adverse health effects towards mechanistic approaches based on human relevant scientific knowledge involving molecular to organism targets and all intermediate levels of complexity. This change of perspective is supported by increased knowledge of molecular mechanisms underlying toxicity, the availability of an abundant array of animal-free test methods, and the expanding work on the description of AOPs, integrated toxicity testing strategies and integrated approaches to toxicity testing and assessment.
The application of these innovative approaches is especially challenging in the area of developmental toxicity, with the developing embryo as its moving target, changing its form, its physiology and its susceptibility to exposures continuously as morphogenesis progresses. The complexity of embryogenesis and its time- and location-specific changes in susceptibility require an integral approach to mechanistic developmental toxicology.
Thus, there is a need for an ontology specific to developmental toxicity that would enable computer-based prediction of which chemicals are likely to induce human developmental toxicity. The ontology should be built by developmental toxicity experts in collaboration with ontology experts. The AOP concept plays a critical role in the ontology by facilitating connections between the chemicals, biological processes, and adverse outcomes.
This report has described some of the principles and approaches feeding into the definition and derivation of a developmental ontology, which could serve as a tool for an integrated assessment of developmental toxicity. Several examples of activities feeding into the development of such an ontology are mentioned, such as the US EPA Virtual Embryo project, the ToxCast database of alternative assays, and the Retinoic Acid Pathway of (dys)morphogenesis.
Combining all existing knowledge into a single developmental ontology will allow the derivation of novel adverse outcome pathways. In addition, it will allow the selection of prioritised biomarkers of adversity throughout the ontology that may be used in efficient integrated approaches of developmental toxicity assessment. More broadly, such an ontology could provide a template for the development of an ontology covering all of toxicity.
 GO: http://amigo.geneontology.org
 KEGG: Kyoto Encyclopedia of Genes and Genomes. KEGG PATHWAY mapping is the process to map molecular datasets, especially large-scale datasets in genomics, transcriptomics, proteomics, and metabolomics, to the KEGG pathway maps for biological interpretation of higher-level systemic functions.