Assessment of Endocrine Disruption Potential of Chemicals Using Combined Quantitative Structure-Activity Relationship Modeling of In Vitro and In Vivo Assays

Abstract

In vitro receptor binding assays for estrogen and androgen systems are widely used for assessing the endocrine disruption potential of chemicals. These assays have generated large amounts of data regularly used for building predictive quantitative structure-activity relationship (QSAR) models. At the same time, in vivo screening assays such as uterotrophic and Hershberger are very valuable because they reflect organ level changes as a result of the interactions of xenobiotics with the endocrine system in physiological conditions. However, such in vivo tests are expensive, time consuming, and require a large number of animals. As a result, very little data are available from these assays, and it is difficult to build useful predictive QSAR models using conventional techniques. In this study, we developed a method to predict in vivo endocrine disruption potential of chemicals using naive Bayes classification models parameterized on the outcomes of QSAR models of in vitro endpoints. The method reduces the need for large amounts of in vivo assay data. In fact, toxicity data of only 25 to 42 compounds were used from uterotrophic, Hershberger agonist, and Hershberger antagonist assays. The model's internal validation performance metrics are in the range of 50%–91% sensitivity, 73%–100% specificity, and 69%–88% accuracy in predicting in vivo outcomes. Balanced accuracies are 87%, 75%, and 70% for the models of uterotrophic, Hershberger agonist, and Hershberger antagonist effects, respectively. On a small external uterotrophic data set of nine compounds with only one negative, the method predicted with 100% accuracy.