A review on advances in chemical compound toxicology prediction

Abstract

The accurate prediction of chemical toxicity remains a major challenge in drug development and chemical safety assessment because conventional approaches can be costly, time-consuming, ethically constrained, and limited in predictive performance. This review examines how artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), is being applied to predictive toxicology. It covers conventional algorithms such as support vector machines (SVMs), decision trees (DTs), and ensemble methods, together with advanced approaches including graph neural networks (GNNs). The review evaluates AI-based methods across five organ-specific toxicity domains—hepatotoxicity, cardiotoxicity, respiratory toxicity, nephrotoxicity, and neurotoxicity—and critically compares molecular representations, data-integration strategies, model architectures, validation protocols, and interpretability approaches. Particular emphasis is placed on data quality, class imbalance, model explainability, applicability domains, external validation, and regulatory acceptance. By integrating recent methodological advances with persistent implementation challenges, this review provides a technical framework for developing more transparent, reliable, and generalizable computational tools for safety assessment in drug discovery and environmental health.