Modeling Developmental Toxicity Hazard Prediction using the ToxCast Library
On this page:
New approach methods (NAMs) leverage chemical-specific pathway information from high-throughput screening (HTS) in vitro assays such as ToxCast/Tox21 to quickly evaluate the human toxicity potential for thousands of chemicals. Here, we present a method for utilizing HTS assays to predict prenatal developmental toxicity. Using a high-confidence reference set (127 compounds) for prenatal developmental toxicity, we trained numerous machine learning algorithms using 50% activity concentrations (AC50s) from the broader ToxCast/Tox21 HTS portfolio (1411 assays) to develop a deterministic developmental toxicity predictive model. Representing a top-down approach, we applied unsupervised filtering criteria and supervised learning models to remove uninformative assays and select for the ToxCast/Tox21 features most predictive of developmental hazard based on the training set of compounds representing clear evidence in the pregnant rat and/or rabbit (positive calls) or no evidence of developmental toxicity (low-effect-level > 1000 mg/kg/d) in the rat and rabbit (negative call). Within the ToxCast/Tox21 portfolio, the Stemina devTOXqp (STM) assay that measures the drop in ornithine-to-cystine in the culture medium of human pluripotent H9 stem cells (hPSCs) was determined to be the most predictive assay. When augmenting STM with 109 additional features following unsupervised and supervised learning, reference set balanced accuracy (BAC) exceeded 77% (63% sensitivity, 90% specificity). We also explored a bottom-up approach which complemented the STM assay with select features expected to be predictive of developmental hazard based on mechanistic relevance: predicted androgen receptor activity (COMPARA), predicted estrogen receptor activity (CERAPP), and a germ layer reporter platform for endoderm differentiation (GLR). Using a Bayesian ensemble model, prediction of the high confidence dataset resulted in a BAC of 75% (77% sensitivity, 74% specificity). While both approaches apply hPSC biology, determination of the complementary pathways for predicting developmental toxicity for chemicals across the broader ToxCast landscape differs. Selection of these additional assays can further be leveraged to maximize sensitivity/specificity depending on decision making requirements. The results from this study highlight the need for a recursive feature addition methodology where mechanistically-indicated assays (method two) are supplemented with a minimal set of predictive HTS assays from the ToxCast library (method one). This abstract does not reflect EPA policy.