Machine Learning-Based Quantitative Structure Activity Relationship Modeling of Repeated Dose Toxicity: A Data-Driven...

In recent years, the rapid increase in the production and environmental release of synthetic organic chemicals has raised serious concerns about their potential adverse effects on human health and the environment. Repeated exposure to such substances can lead to significant toxicological effects, underscoring the importance of early and reliable hazard assessment. However, experimental determination of repeated-dose toxicity (RDT) is costly, time-consuming, and constrained by ethical considerations. In this study, we developed various classification-based predictive models to evaluate the subchronic RDT potential of chemicals after oral exposure. We compiled data from eChemPortal and J-CHECK databases. The data set contains two study-derived effect levels: NOAEL (no observed adverse effect level) and LOAEL (lowest observed adverse effect level), for which separate models have been developed. A key strength of this data set is that all studies followed standardized OECD test guidelines (407, 408, and 422) and were conducted under good laboratory practice (GLP) conditions, ensuring regulatory relevance and high data reliability. Multiple machine learning algorithms were systematically evaluated, and the best models were selected using a multicriteria analysis based on the sum of ranking differences (SRD) technique. The final selected models achieved accuracies on the training sets ranging from 0.665 to 0.902, while the test sets showed accuracies ranging from 0.642 to 0.682. We also conducted a substructure analysis to identify the key substructures involved in the toxicity. This analysis revealed eight structural motifs, with chlorine- and amine-group-containing aromatic systems being particularly significant. The final developed models were experimentally validated using chemical substances provided by Global Product Compliance (GPC) Europe AB. Additionally, the models were applied to the Pesticides Properties DataBase (PPDB) to screen for pesticides with potential toxicity upon repeated exposure. To facilitate accessibility and regulatory application, the final models have been implemented in both a Python-based tool and a web application. Scientific contribution: this study presents predictive models as alternatives to traditional animal testing for assessing the subchronic oral repeated-dose toxicity (RDT) of chemicals. Our models demonstrate strong statistical performance, indicating their suitability for further application, as supported by experimental validation. These models could be used for preliminary hazard screening or weight-of-evidence evaluations. An additional advantage is that these models were developed using data that were tested in accordance with internationally harmonized test protocols, thereby enhancing their acceptance for regulatory decision-making.