Sustainable development goals, including SDG-6 (clean water and sanitation), SDG-3 (good health and well-being), and SDG-12 (responsible consumption and production), are the primary emphasis of this research, which aims to improve chemical safety through accurate and reliable prediction of acute oral toxicity. We need to find more effective ways to conduct toxicity evaluations because the old ways are so expensive and take so long. Finding the best way to build accurate and dependable toxicity prediction models is the primary goal of this research. Using a comprehensive dataset from the Interagency Center for the Evaluation of Alternative Toxicological Methods/National Toxicology Program (NTP) and the National Center for Competency Testing/Environmental Protection. For the purpose of agency, we evaluated the severe oral harmfulness prediction capabilities of random forests (RF), a strong ensemble approach. With just 8% of medications being extremely toxic and 92% being mildly harmful, there is a clear disparity between the two classes. To fix this mismatch, we used cost-sensitive learning (CSL) and data resampling methods like under-sampling and over-sampling. As input characteristics, the model underwent normalization, validation, and feature selection. The rational discovery kit was utilized to build a broad set of two-dimensional chemical descriptors. Bayesian optimization and cross validation were used for hyper-parameter tuning, and random forests were tested against the generalized linear model, artificial neural networks, gradient boosting, extreme gradient boosting, and extreme boosting. The RF models, especially those employing under sampling and CSL, exhibited superior performance on an independent test set, achieving an accuracy of 0.88, specification of 0.88, sensitivity of 0.86, and an AUC of 0.92 for the receiver operating characteristic curve. Chemical descriptors that are most important, according to feature importance analysis, are a molecule’s quantum numbers and its Van der Waals surface area. With the use of random forest predictions, a surrogate decision tree was able to achieve an AUC of 0.948. When resampling and cost-sensitive learning were used to solve class imbalance, random forest models were able to effectively predict acute oral toxicity. Using methods from explainable AI, such as local interpretable model-agnostic explanations, surrogate decision tree analysis, and permutation feature importance, the study identified key chemical properties that induce toxicity. An important step in improving chemical safety and supporting sustainable development goals, this research also improves the interpretability of models.