Deep Learning in Environmental Toxicology: Current Progress and Open Challenges

Ubiquitous chemicals in the environment may pose a threat to human health and the ecosystem, so comprehensive toxicity information must be obtained. Due to the inability of traditional experimental methods to meet the needs of toxicity testing of a large number of chemicals, in vivo and in vitro assays have been shifted to a new paradigm, computer-assisted virtual screening. However, the commonly used virtual screening techniques, including read-across and machine learning-based quantitative structure–activity relationship (QSAR), have limitations in assessing complicated, high-dimensional, and multimodal bioactivity data. In these cases, deep learning (DL) has emerged as a desirable solution for the application of QSARs in toxicity prediction. Therefore, this paper introduces and discusses (a) architectures of six commonly used DL algorithms (fully connected neural network, convolutional neural network, recurrent neural network, long short-term memory, graph neural network, and generative adversarial network), (b) the application scenarios of six DL algorithms, e.g., toxicity prediction and data generation, and (c) challenges and future trends of DLs in toxicity prediction. We believe that by consolidating toxicological mechanisms and DL algorithms, this survey can help readers to build prediction models with excellent performance while promoting further discussions of the fusion of environmental toxicology and DL.