|
Abstract:
|
Models capable of predicting dose-response curves have several potential applications in toxicological risk assessment. For instance, such models could be used to examine relationships between different assays or predict toxicity based on chemical properties. In toxicogenomic studies, one might be interested in a model that can predict the impact of chemical exposure on certain genes based on their function. However, constructing prediction models can be challenging because the shape of outcome distributions often change with dose thus violating assumptions of standard parametric models. Thus, motivated by data from toxicogenomic studies, we propose a Bayesian nonparametric approach for predicting the shape of a dose-response function. A stick-breaking process is used to relax distribution assumptions and cluster dose-response curves from different genes based on shape. Cluster-specific curves are modeled using regression splines subject to an umbrella ordering constraint. A logistic regression model is used for the "stick-breaks" to predict dose-response based gene-specific covariates (e.g., gene function). The method is applied to a recent National Toxicology Program study.
|
