Identification of Structural Alerts by Machine Learning and Their Applications in Toxicology

As a simple and intuitive method to predict chemical toxicity, structural alerts have been widely used in toxicology research and drug discovery. In the past decade, methods to identify structural alerts have made considerable progress, which allows for the automatic identification of structural alerts and achieves a high accuracy in toxicity prediction. However, these computational methods focus on prediction accuracy of toxicity rather than the mechanism of toxicity, resulting in a series of controversial structural alerts. Many studies demonstrate that the integration of knowledge-based structural alerts and machine learning models can greatly improve the interpretability, credibility, and impact of toxicity prediction results. In this chapter, we describe the development of approaches for identification of structural alerts, i.e., from expert systems to computational approaches, and introduced the applications of structural alerts in toxicology. In addition to toxicity prediction, structural alerts can also be used to explain quantitative structure–activity relationship models, to optimize molecules and to explore new toxicity mechanisms. Nevertheless, some challenges remain open problems to be solved, such as how to address the co-existence of several structural alerts in one molecule and how to explore new toxicity from data-driven structural alerts.