Mining environmental chemicals with boosted trees
Natalia Khuri · 2020
In response to scientific reports and legislative actions, the United States Environmental Protection Agency (EPA) launched an Endocrine Disruptor Screening Program (EDSP). The overarching aim of this program is to develop novel analytical and computational tools and methods for the detection of environmental chemicals causing aberrations in estrogen, androgen, or thyroid hormone systems. As a result of this program, over 700 diverse environmental chemicals have been tested in in vitro and in vivo assays and deposited into public databases. In this work, machine learning classifiers were developed to predict putative disruptors of the human sodium/iodide symporter that plays a crucial role in the biosynthesis of thyroid hormones. Two powerful ensemble algorithms, Random forest and eXtreme Gradient Boosting Tree, were trained with EDSP experimental data and evaluated by repeated cross-validation as well as by retrospective validation. Within its applicability domain, Boosted Tree classifier achieved high performance and discriminated between inhibitors and noninhibitors with an accuracy of 87%, precision of 85% and recall of 63%. Additionally, 98 inhibitors were also predicted among 1,741 human endogenous and exogenous metabolites, including approved oral drugs. Further experimental studies will be needed to validate these predictions and elucidate the mechanism of interactions.