ExNCDIA: an explainable ensemble learning framework for predicting non-chemotherapy drug-induced agranulocytosis and interpreting structural associations

Non-chemotherapy drug-induced agranulocytosis (NCDIA) is a rare but life-threatening idiosyncratic toxicity whose mechanisms remain largely unresolved; for a subset of drugs, however, the risk has been linked to chemical structure through bioactivation, whereby reactive intermediates injure neutrophils and myeloid progenitors, and haptenate proteins to trigger immune-mediated destruction. Yet the structural features underlying this liability remain poorly defined, and interpretable models for flagging it during design are scarce. Here we report ExNCDIA, an explainable machine learning framework that predicts NCDIA liability directly from chemical structure. From a curated set of 906 drugs (371 NCDIA-positive and 535 NCDIA-negative), we benchmarked complementary open-source representations and imbalance-aware ensemble classifiers, and retained two balanced random forest models-one trained on RDKit molecular descriptors refined by genetic-algorithm feature selection (test ROC-AUC 0.850, MCC 0.591), the other on MACCS substructure keys (test ROC-AUC 0.847, MCC 0.518). Both showed balanced discrimination in training-set cross-validation and on the held-out test set. Crucially, Shapley additive explanations (SHAP) applied to these two non-overlapping representations converged on the same medicinal-chemistry signal: NCDIA risk rises with sulfur-containing pharmacophores, multiple aromatic rings, and nitrogen-bearing/arylamine cores-motifs prone to cytochrome P450- and myeloperoxidase-mediated oxidation to reactive metabolites-and falls with aliphatic, saturated-oxygenated scaffolds lacking such bioactivatable groups. Delivered as a freely accessible web tool, ExNCDIA couples quantitative prediction with interpretable structural associations, supporting early candidate prioritization and post-marketing risk assessment.