Metal oxide nanoparticles (MONPs) are increasingly used in biomedical applications due to their small size, high surface area, and tunable chemistries, yet these same characteristics raise safety concerns. MONPs can interact intimately with cells, causing toxicity through mechanisms like ion release, reactive oxygen species (ROS) generation, membrane damage, and cellular dysfunction. To enable safer development of MONPs, we created interpretable predictive models linking their physicochemical and exposure properties to cytotoxic outcomes. Utilizing the “Data Curation to Develop Machine Learning Models for Assessing the Toxicity of Nanoparticles” dataset curated from around 140 peer-reviewed publications we engineered a robust feature set and trained classification and regression models. Our two-stage workflow classifies materials into toxicity classes based on their features, which then inform regression models predicting biological outcomes. Using LIME plots for model interpretation, we found that core size was the strongest negative contributor to toxicity, followed by exposure dosage and time, while zeta potential, ion release, and aggregation state had lesser impacts. These results align with known mechanisms of MONP toxicity and highlight the importance of exposure parameters. Our study illustrates how interpretable machine learning can accelerate the design of safer MONPs, reducing reliance on extensive in-vitro screening while improving mechanistic understanding.