(618b) Computational Characterization of the Estrogenic Activity of Environmental Chemicals Using High-Throughput Microscopy and Machine Learning

The goal of this study is to develop a machine learning framework that can classify the estrogenic activity of endocrine disrupting chemicals (EDCs) and analyze the experimental feature subsets that have the most impact on this prediction. The endocrine system is responsible for regulating a variety of physiological processes in living organisms, and any disruption caused by chemical exposures can have long-term impacts on human and wildlife health [1]. Machine learning approaches possess an untapped opportunity for a facilitated toxicity evaluation of environmental chemicals by learning the patterns in the biological activity data, which can assist in predicting the potential consequences of exposure to these compounds [2].

In this work, we analyze 60 reference EDCs with known estrogen receptor-a activities using an engineered biosensor and collect high throughput/high content microscopy data with mineable size, shape, and intensity attributes that relate to the various mechanistic endpoints on the estrogen receptor signaling pathway [3]. We formulate the interaction of EDCs with the estrogen receptor ligand binding domain (i.e., agonist or antagonist) as a nonlinear classification problem. We train, validate, and test Support Vector Machines (SVM), Random Forests (RFs), and Artificial Neural Networks (ANN) for mapping the separation between agonist and antagonist chemicals with implicit and explicit mathematical models. By following our previously studied preprocessing routines [4,5], we demonstrate the effect of data preparation on the classification model performance. We use clustering analysis to identify the most promising feature subsets using the Fowlkes–Mallows index and perform principal component analysis to minimize experimental noise. Finally, the predictive performance of the three classifiers is evaluated on hold-out chemicals (blind testing), and the classification accuracy across the selected feature groups is quantified. The results show that our ANN classifier with trained with benchmark agonist and antagonist chemicals achieves 98.41% blind testing accuracy, where this number drops to 96.36% for the SVM and 94.36% for the RF model.

References

1. Schug, T.T., Janesick, A., Blumberg, B. and Heindel, J.J., 2011. Endocrine disrupting chemicals and disease susceptibility. The Journal of steroid biochemistry and molecular biology, 127(3-5), pp.204-215.

2. Wang, M.W., Goodman, J.M. and Allen, T.E., 2020. Machine learning in predictive toxicology: recent applications and future directions for classification models. Chemical research in toxicology, 34(2), pp.217-239.

3. Ashcroft, F. J., Newberg, J. Y., Jones, E. D., Mikic, I., & Mancini, M. a. (2011). High content imaging-based assay to classify estrogen receptor-α ligands based on defined mechanistic outcomes. Gene, 477(1–2), 42–52.

4. Mukherjee, R., Beykal, B., Szafran, A.T., Onel, M., Stossi, F., Mancini, M.G., Lloyd, D., Wright, F.A., Zhou, L., Mancini, M.A. and Pistikopoulos, E.N., 2020. Classification of estrogenic compounds by coupling high content analysis and machine learning algorithms. PLoS computational biology, 16(9), p.e1008191

5. Aghayev, Z., Szafran, A.T., Tran, A., Ganesh, H.S., Stossi, F., Zhou, L., Mancini, M.A., Pistikopoulos, E.N., Beykal, B., 2023, Machine learning methods for endocrine disrupting potential identification based on single-cell data, Chemical Engineering Science (Under Review).