Tuesday, 3/17/20 @ 9:00 AM – 10:45 AM | Exhibit Hall – Poster Board No. P383
Predicting Developmental Toxicity Potential Using Pluripotent Embryonic Stem Cell Assays and the ToxCast Library
T.J. Zurlinden1, N.C. Kleinstreuer2, A. Lumen3, M. Xia4, N.C. Baker5, and T.B. Knudsen1
1 US EPA, Research Triangle Park, NC USA
2 NIEHS / NICEATM, Research Triangle Park, NC USA
3 US FDA / NCTR, Jefferson, AR USA
4 NIH / NCATS, Bethesda, MD USA
5 Leidos, Research Triangle Park, NC USA
New approach methodologies are being explored for their ability to quickly evaluate the human toxicity potential of chemicals with less reliance on animal testing using ToxCast / Tox21 generated in vitro data on thousands of chemicals utilizing high-throughput screening (HTS) assays. Leveraging a high-confidence reference set (127 compounds) of prenatal developmental toxicity, we trained numerous machine learning algorithms using AC50s from the broader ToxCast / Tox21 HTS portfolio (1411 assay features) to develop a deterministic predictive model for human developmental toxicity.
Within the ToxCast / Tox21 portfolio is the Stemina devTOXqp assay that measures a critical drop in the ornithine / cystine (O/C) ratio in the culture medium of H9 human embryonic stem cells (hESCs) maintained pluripotent for a 3-day exposure window to provide a predictive biomarker for human developmental toxicity potential.
We obtained a positive signal on 183 (17%) of 1062 ToxCast chemicals tested to date. Balanced accuracy (BAC) of 62% was shown for 432 chemicals having any developmental toxicity evidence from studies in pregnant rats or rabbits (Toxicological Reference Database: ToxRefDB); however, BAC exceeded 77% (63% sensitivity, 90% specificity) for a subset of 127 compounds with strong evidence of developmental toxicity or non-toxicity.
The winning model was L2-penalized logistic regression, with an overall 5-fold cross validation BAC of 81% (70% sensitivity, 92% specificity) against the 127-compound training set. Evaluation of the log-odds coefficients from the logistic regression model selected the hESC devTOXqp O/C ratio as the most predictive ToxCast assay feature.
Finally, we trained a Bayesian logistic regression model with L2-regularization on the entire 127-compound training set to develop a probabilistic model for predicting developmental toxicity. Using ToxCast bioactivity profiles, the model predicted developmental toxicity potential of the remaining 305 chemicals with in vivo prenatal developmental toxicity evidence from ToxRefDB and the literature, allowing for external validation of these predictions. Overall, this model leverages hESC biology and complementary pathways to make probabilistic predictions of developmental toxicity potential for chemicals across the broader ToxCast landscape.
This abstract does not reflect US EPA policy.