Computational Deep Assessment of Acute Aquatic Toxicity Using Specific Machine Learning Approaches In MATLAB Environment

The accelerating production and dissemination of synthetic chemicals have intensified the demand for scalable and ethically responsible toxicity assessment strategies. Conventional experimental testing, particularly animal-based acute toxicity assays, is increasingly constrained by financial cost, limited throughput, and regulatory pressure to reduce animal use. In response, computational toxicology has gained prominence as a data-driven alternative capable of supporting early-stage hazard screening and regulatory decision-making. This study proposes a machine-learning-centered framework for the prediction of acute aquatic toxicity, focusing on fish lethality endpoints quantified by some metal compounds such as Cr, Mn, Fe, Co, Ni, CU, Zn, As, Cd and Pb in a specific region of Gülşehir, Türkiye with using the fish species of Cyprinus carpio (5 male and 5 female data). By integrating molecular structure representations, physicochemical attributes, and statistical learning algorithms, the framework captures non-linear and high-order interactions that are poorly represented by traditional quantitative structure–activity relationship models. Multiple supervised learning paradigms— including ensemble tree methods, kernel-based classifiers, and neural architectures—are examined in both regression and categorical toxicity classification contexts. Model reliability is evaluated using stratified validation schemes, external testing, and applicability-domain constraints to ensure transferability across chemically heterogeneous datasets obtained by us for this study. The findings indicate that machine learning models provide a substantial improvement in predictive consistency and generalization when compared with linear and semi-empirical approaches. According to the results, 4 Machine Learning (ML) methods of Gradient Boost, Random Forest, SVR (Radial Basis Function) and Multilayer Perceptron (MLP) were used diversily and the highest performance metrics were obtained for Gradient Boosting method of %96 and Random Forest method for %94, approximately. In addition, interpretability techniques are incorporated to elucidate structurally relevant contributors to toxicity, thereby enhancing scientific transparency and regulatory acceptability. Collectively, the results reinforce the role of machine learning as a practical and scientifically robust component of nextgeneration toxicity prediction pipelines, with direct implications for chemical prioritization and animal-testing reduction.