Nonparametric Bayesian Joint Modeling of High-Throughput and Low-Throughput Genotoxicity Data

• Activity Number: 461 - Bayesian Statistical Methods for High-Throughput Toxicity Testing and Risk Assessment
• Type: Topic Contributed
• Date/Time: Wednesday, July 31, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Section on Risk Analysis
• Author(s): Michael Pennell* and Matthew W Wheeler
• Companies: Ohio State University and CDC/NIOSH
• Keywords: Count Data; Dose-Response; Function-on-scalar Regression; Nonparametric Regression; Stick-Breaking Process; Toxicological Risk Assessment

Abstract

The goal of the EPA ToxCast program is to develop in vitro high-throughput screening techniques that can prioritize thousands of environmental chemicals requiring toxicity testing. While the program has been successful in developing assays relevant to in vivo toxicity, their relationship to traditional low throughput assays is uncertain. For example, it is unknown if ToxCast assays measuring cytotoxicity, tumor suppression response, and other stress responses predict results from the Ames Assay, an in vitro assay that is the gold standard for predicting genotoxicity. To make this connection, we develop a Bayesian nonparametric approach for joint modeling Ames and ToxCast assay data. Count data from the Ames assay are modeled using a Logistic Stick-Breaking Process (LSBP) whose locations differ with dose according to a nonparametric regression function with an umbrella ordering. The weights of the LSBP are modeled using a function-on-scalar regression of over 250 ToxCast assays. The method is used to simultaneously model genotoxicity data on 617 chemicals. The predictive ability of the model is evaluated using a hold-out sample of 155 chemicals.