Screening and prediction of potential androgen receptor agonists in indoor dust

Androgen receptor (AR) agonists in indoor dust have raised significant concerns due to their endocrine-disrupting properties. However, the screening and prediction of AR agonists remain challenging. In this study, 2092 chemicals in indoor dust were identified by gas chromatography/quadrupole time-of-flight mass spectrometry-based non-targeted analysis. Known AR agonists among those chemicals, such as benzo[a]pyrene, benzo[b]fluoranthene and butyl acrylate, were screened using the United States' Tox21 database, and potential AR agonists not contained in the Tox21 database were predicted by quantitative structure-activity relationship (QSAR) models based on six machine learning algorithms (random forest, extremely randomised trees, support vector machine, light gradient boosting machine, extreme gradient boosting machine and soft voting ensemble model). Based on the performances of the models in five-fold cross-validation on the training set and test set, the soft voting ensemble model constructed using the original training set was selected as the optimal model. Four compounds, namely 21-hydroxypregnenolone, 7-oxycholesterol acetate, cholesterol heptanoate and tripropylene glycol diacrylate, were successfully predicted to be potential AR agonists and should be paid more attention in future work.