Interpretable QSAR modelling for immunotoxicity prediction using enhanced fingerprint and SHAP-based feature selection

Accurate prediction of immunotoxic effects is essential for chemical safety evaluation and drug development. However, existing methodologies are limited by the scarcity of in vitro data and the inherent complexity of immune responses. This study introduces an interpretable quantitative structure-activity relationship (QSAR)-based modelling framework aimed at assessing immunosuppressive toxicity utilizing IC₅₀ data obtained from three human immune cell lines: Jurkat, peripheral blood mononuclear cells (PBMC) and THP-1. Three tree-based machine learning algorithms, in conjunction with robust feature selection techniques, were employed to identify critical molecular determinants associated with immunosuppressive effects. The implementation of SHapley Additive exPlanations (SHAP) enhanced model interpretability and facilitated the extraction of potential structural alerts, thereby providing mechanistic insights into immunotoxicity pathways. Our findings indicate that the integration of immune cell-specific experimental data with interpretable modelling approaches significantly enhances the reliability of immunotoxicity predictions. This research establishes a scientifically grounded framework that not only supports the early identification of immunotoxic chemicals but also promotes safer chemical design and informed decision-making in drug development and toxicological risk assessment.