Shaping the Future of Predictive Toxicology with Machine and Deep Learning: A Comprehensive Review

Abstract

One important part of drug development and chemical safety evaluations is predicting toxicity. Most conventional methods for toxicity assessment rely heavily on animal experiments, are time-consuming, costly, ethically dubious, and not scalable. In addition, the expanding chemical space and limitations of in vivo assays have necessitated the development of more rapid, reliable, and high-throughput computational alternatives. AI, particularly the application of machine learning (ML), deep learning (DL), and related AI technologies, is transforming predictive toxicology. The application of Generative AI in molecular generation, toxicity-aware compound design, data augmentation, and virtual screening presents new avenues for advancing drug discovery more safely and efficiently. This review discusses recent progress in ML, DL, and AI-based strategies for toxicity prediction, assesses their utility across major toxicity endpoints, and highlights the challenges and future research opportunities in predictive toxicology. Molecular descriptors, fingerprints, SMILES strings, and graph-based molecular structures are examined across acute, chronic, hepatotoxicity, cardiotoxicity, nephrotoxicity, neurotoxicity, and carcinogenicity. Algorithms such as support vector machines, random forests, convolutional neural networks, recurrent neural networks, graph neural networks, generative adversarial networks, and multitask learning frameworks are critically examined, alongside publicly available toxicity datasets and field-specific computational tools. Random forests, support vector machines, graph neural networks, and deep learning architectures currently show the strongest predictive performance across these endpoints. Nevertheless, the most frequent issues plaguing AI-based toxicity models include data imbalance, poor interpretability, limited reproducibility, and insufficient generalization across different domains. Future advancements in explainable AI, multimodal data integration, transfer learning, generative AI, and high-throughput screening may yield better-performing, interpretable, and regulatory-robust toxicity models.