This presentation will provide an accessible overview of how quantitative structure-activity/toxicity relationship (QSAR/QSTR) modeling and explainable machine learning can support early drug-safety assessment. The talk will introduce the main steps of an interpretable computational workflow, including molecular descriptor generation, leakage-controlled model development, independent and external validation, applicability-domain assessment, and SHAP-based explanation. Drug-induced neutropenia will be used as a case study to illustrate how molecular structure can be translated into predictive and interpretable safety signals. Results from a descriptor-guided Random Forest model will be presented to demonstrate how such approaches may help identify structural patterns associated with neutropenia risk and support toxicity-aware prioritization. The emphasis will be on the broader value, limitations, and practical interpretation of explainable QSAR models rather than on a single modeling result.