AOP-Based Machine Learning for Toxicity Prediction

On a daily basis, people will be exposed to various compounds, many of which potentially endanger human life and health. Therefore, there is a driving imperative to evaluate the toxic effects of these compounds. However, the cost and time required to obtain toxicity data through traditional in vivo experimental methods are high. The promise of obtaining toxicity data through virtual screening, especially machine learning (ML) algorithms, has attracted increasing attention. However, ML-based toxicity prediction faces many challenges. ML algorithms have fundamental limitations, interpretability can be limited with “black box” approaches, and forecast accuracy leaves room for improvement. In these cases, the adverse outcome pathway (AOP) provides a theoretical framework and predictive feasibility for virtual screening based on ML algorithms. Integrating the concept of the AOP framework into ML-based modeling can provide meaningful biological interpretations for predictive models. To date, many have successfully confirmed the feasibility of using the AOP framework to construct virtual screening models. In this chapter, we will discuss the problems encountered by ML techniques in toxicity prediction. Additionally, we will detail the advantages of ML techniques based on the AOP framework for toxicity prediction and offer some relevant examples. ML algorithms based on the AOP framework promise to deliver breakthrough progress in toxicity prediction in the near future.