Improving inferences on health effects of chemical exposures

Adverse effects of environmental contaminants on human health are a major public health concern. We are all exposed to a complex mixture of different chemical contaminants through the air we breathe, the water we drink, the food we eat, and the products we use. As new industrial products are produced, leading to new direct and indirect exposures, there is a pressing need for new tools for assessing the adverse health effects in humans associated with exposure to chemical mixtures. Challenges include huge numbers of different possible mixtures, the curse of dimensionality in multivariate nonparametric regression and moderate to high correlation in different exposures. Building on compelling preliminary results from a highly successful NIEHS PRIME program R01, we develop a transformative statistical toolbox for inferences on health effects of chemical exposures, both in the high throughput screening context and for better disentangling health effects of chemical mixtures in epidemiology studies. The research proceeds through the following Aims. Aim 1 develops methods for inferring synergistic and antagonistic interactions from epidemiologic data, including for data collected longitudinally motivated by studies of exposure effects on childhood neurodevelopment. We improve substantially over current nonparametric regression approaches in interpretability and power to detect interactions; synergistic interactions in which chemicals amplify each other’s effects are particularly important. Aim 2 develops clustering methods to improve understanding of variation in exposure in relation to health. These methods will have broad impact in dramatically improving practical performance over current model- based clustering approaches. In addition, easily interpretable results are provided, adding additional insights over state-of-the-art regression-based methods. Aim 3 develops new methods for inferring relationships between chemical molecular structure and biologic activity. Given the sheer number of chemicals lacking any direct in vivo or in vitro data, it becomes crucial to use molecular structure to predict biologic activity. Leveraging on ToxCast/Tox21 and other data sources, we develop improved statistical models for relating chemical structure to activity, for inferring low-dimensional summaries of chemical activity based on molecular structure, and for optimally choosing the next chemicals to be tested. These methods can be used to predict effects of chemicals lacking any direct in vivo or in vitro data through targeted borrowing of information across related chemicals in the database. Aim 4 develops user-friendly and reproducible software, while using the methods to thoroughly analyze data from the motivating epidemiology studies, with a particular focus on the Mount Sinai Children’s Environmental Health Study and the UNC Early Life Factors Study, which both focus on assessing exposure effects on neurodevelopment in early childhood. We expect our methods to lead to important new findings.