Pulmonary toxicity caused by inhaled nanoparticles is an endpoint of high concern yet there are major ethical and financial limitations to in vivo bioassay. Development of in silico models can fill gaps in pulmonary toxicity data, avoid testing of each new nanoparticle formulation from scratch and facilitate safe nanotechnology. In this study, machine learning models were built to relate metal oxide nanoparticle (MeONPs)' physicochemical properties to their toxicity in an alveolar-macrophage-like cell (THP-1), which is the main route to remove inhaled insoluble fine particulates. The endpoints to be predicted were inflammatory potential based on a pro-inflammatory cytokine (IL-1β) release and cytotoxicity based on cell viability. Dataset for each endpoint consists of 240 toxicity data (30 MeONPs tested at 8 serial dilutions), with 9 physicochemical properties and 59 quantum-mechanical attributes (input parameters). The in vitro toxicity data were validated in mouse lungs by oropharyngeal instillation of six representative MeONPs. Inflammatory potential and cytotoxicity were correctly predicted with predictive accuracy (ACC) exceeding 90% and further validated experimentally with ACC reaching 86% and 75%, respectively. The structure–activity relationship extrapolation indicated that electronegativity, ζ-potential, and cation charge were key properties responsible for inflammatory effects of MeONPs; while size, dissolution and electronegativity of MeONPs could be used to determine their cytotoxicity. DFT computations further revealed the underlying mechanisms: MeONPs with lower metal electronegativity and positive ζ-potential were more likely to cause lysosomal damage and substantial inflammation or cytotoxicity. The insights derived are useful to consider when developing future (non-)testing approaches to address regulatory purposes.