An image based approach for predicting the effects of endocrine disrupting chemicals on human health using deep learning

In recent years, deep neural networks, especially those exhibiting synergistic properties, have been at the cutting edge of image processing, producing very good results. So far, they have been able to successfully address issues of classification and recognition of objects depicted on images. In this paper, a novel idea is presented, where images of chemical structures are used as input information in deep learning neural network architectures aiming at the generation of Quantitative Structure Activity Relationship (QSAR) models, i.e. models that predict properties, activities or adverse effects of chemicals. The proposed method was applied to a case study of particular interest, which is the prediction of endocrine disrupting potential of chemicals. Two different deep learning architectures were applied. The produced ImageNet model proved successful, in terms of accuracy, performance and robustness on training and validation sets. The new approach is proposed to the community as an alternative or complementary method to current practices in QSAR modelling, which can automate and improve the creation of predictive models.