Ciallella, Heather L.. Predicting developmental and reproductive toxicity using artificial intelligence and high-throughput screening data. Retrieved from https://doi.org/doi:10.7282/t3-70wa-0v48
DescriptionUnder current regulatory guidelines for identifying hazards, new and existing chemicals must undergo expensive, time-consuming, and ethically questionable animal testing to evaluate their developmental and reproductive toxicity (DART) potential. The in vitro assay data generated by high-throughput screening programs has emerged as promising alternatives to animal testing. This dissertation presents new methods for predicting DART-related endpoints. Estrogenic chemicals and pathways have well-studied DART effects, providing a strong foundation for developing novel computational models. First, over 7500 chemicals tested in the EPA’s Toxicity Forecaster (ToxCast) initiative were used to develop 276 machine and deep learning models to predict nuclear estrogen receptor (ERα/ERβ) activity (AUC=0.56-0.86). Next, this study’s workflow was made automatic and freely available for public users with limited programming knowledge to develop models for other in vitro assays. Then, a knowledge-based deep neural network (k-DNN) approach was designed to accurately identify estrogen mimetics by virtually simulating the genomic ERα/ERβ pathway (AUC=0.864-0.927). Finally, a comprehensive prenatal developmental toxicity database (n=1244 chemicals) was compiled. Relevant PubChem and ToxCast assay data were grouped based on their chemical-mechanistic relationships, revealing two large clusters with high predictivity for new prenatal developmental toxicants (PPV=72.4% and 77.3%). The three methods presented here are potentially universal strategies for leveraging publicly available data to predict chemical hazards.