Bisphenol S (BPS) is widely used as a replacement for bisphenol A, yet accumulating evidence suggests that it has comparable endocrine and cardiovascular toxicity. Here, we investigated whether prolonged low-dose BPS exposure induces subtle but classifiable phenotypic alterations in human coronary artery endothelial cells (HCAEC), using an end-to-end experimental and ML pipeline that spans cell culture, high-content imaging, feature extraction, and robust classification. Cells were exposed to 0.1 µM BPS for 96 h and profiled using a cell painting assay and high-content microscopy. Image segmentation yielded ~2500 quantitative features per cell across four compartments—Membrane, Cytoplasm, Ring region (i.e., perinuclear region), and Nucleus—for multiple fluorophores. We systematically compared different classifiers (Random Forest, XGBoost, LASSO logistic regressor) using feature selection (MRMR, ReliefF, LASSO) or transformation-based dimensionality reduction (PCA, autoencoders). Tree-based ensembles robustly handled high-dimensional inputs, with XGBoost combined with ReliefF-selected features achieving the best performance. The most informative descriptors predominantly mapped to mitochondrial and nuclear channels, indicating early alterations in mitochondrial organisation and chromatin-related features. These findings show that chronic low-dose BPS exposure elicits a distinct endothelial phenotype, consistent with early endothelial dysfunction, and demonstrate that integrating high-content imaging with machine learning provides a sensitive, scalable framework for vascular toxicity assessment of environmental contaminants.
Ferariu et al. (Fri,) studied this question.