Introduction/Objective: High rate of attrition still inhibits drug discovery and development, with toxicity accounting for one of the primary causes of failure in preclinical and clinical development. This review summarizes Machine Learning (ML), Deep Learning (DL), and emerging post-deep learning strategies in drug discovery and environmental safety. Methods: Following PRISMA guidelines, a systematic search was conducted across PubMed, Web of Science, Scopus, and ScienceDirect for the years 2015–2025, yielding 1,020 articles. Additional records were obtained from Google Scholar and the reference lists of about 60 articles. The studies were included when they used ML/DL to predict toxicity, provided quantitative measures of performance (e.g., accuracy, AUC, F1-score), or when they described the predictive tools and platforms. Eligibility criteria were: the study was entirely experimental toxicology with no computational modeling of the study, lacked an adequate description of the methodology, or was in a non- English language. The last study count of the paper is 50 articles. Results: DL models like convolutional and graph neural networks are more effective in cases when the size of the datasets is large. Recent methods that address the problem of data scarcity are property augmentation, transfer learning, and semi-supervised learning. A number of web-based applications (e.g., ADMETlab 3.0, admetSAR 3.0, ProTox 3.0) have been published that allow using multi-endpoint prediction with different measures of accuracy and interpretability. Discussion: Traditional ML methods, particularly support vector machines and random forests, remain valuable for smaller datasets due to their robustness and interpretability. However, the adoption of deep learning architectures, such as convolutional and graph neural networks, has markedly improved predictive accuracy when applied to large and complex datasets. Conclusions: Data-driven methods have significantly advanced toxicity prediction, offering faster and more cost-effective tools compared with traditional assays. However, the field still faces challenges related to limited datasets, variable data quality, and a lack of mechanistic interpretability.
Al-Hussaniy et al. (Tue,) studied this question.