周期表に基づく解釈可能な機械学習と記述子を用いた材料横断的なナノ毒性予測

The biological effects of engineered nanomaterials are largely determined by their intrinsic physicochemical properties; however, predicting these effects across materials with different core compositions remains a significant challenge. In this study, we developed an interpretable machine learning framework to enable cross-material nanotoxicity prediction using descriptors based on the periodic table. A curated data set comprising 1206 entries of metal oxide nanoparticles with cytotoxicity endpoints was constructed from high-quality literature sources. We evaluated multiple machine learning models to systematically analyze their performance using both within-material and cross-material validation strategies. Initial models based solely on experimental conditions and global nanomaterial properties showed high accuracy in within-material prediction but failed to generalize to unseen materials. To overcome this limitation, we introduced elemental-level descriptors, including electronegativity, ionization energy, atomic radius, and oxidation state. Incorporating these fundamental features significantly improved cross-material prediction performance, with coefficient of determination (R²) values from approximately 0.35 to 0.65 for unseen nanomaterials such as CuO, ZnO, MgO, and Cu₂O. Further experimental validation using NiO and Cr₂O₃ nanoparticles on A549 and HUVEC cell lines confirmed the reliability of model predictions and highlighted material- and cell-type-specific toxic responses. Model interpretation and oxidative stress assay provided mechanistic insights into the cytotoxicity induced by the metal oxide nanoparticles. This study demonstrates the feasibility of cross-material toxicity prediction and establishes a scalable and interpretable computational framework for safer nanomaterial design and risk assessment.