Study-Grouped Predictive Toxicology of In Vitro Polystyrene Micro- and Nanoplastic Cytotoxicity: Geometry-Derived Administered-Dose Metrics and a Nanoscale-Restricted Sensitivity Analysis

Polystyrene (PS) micro- and nanoplastics show heterogeneous in vitro cytotoxicity across particle characteristics and exposure conditions. We examined whether geometry-derived administered-dose metrics improve out-of-study prediction of in vitro cytotoxicity and whether ≤100 nm restriction improves performance independently of study composition. From the 2443 literature-derived observations across 98 mammalian in vitro studies, 1701 observations from 82 studies formed the primary mass-concentration dataset and 995 from 60 studies formed the ≤100 nm subset. Random forests were evaluated using nested study-grouped cross-validation with inverse-study-frequency weights. A matched dataset of 141 observations from six studies enabled leave-one-article-out comparisons of mass, particle-number, and geometric surface-area metrics. In the nanoscale subset, observation-level mean absolute error decreased from 15.36 to 14.37 percentage points and R2 increased from 0.283 to 0.447, whereas study-balanced mean absolute error did not improve (15.37 versus 15.60). In a composition-controlled 22-study analysis evaluated on identical nanoscale test observations, nanoscale-only training showed no clear advantage. Ridge regression favored mass concentration, while random forest showed a small, uncertain particle-number advantage. Geometry-derived administered-dose metrics offered no consistent model-independent advantage. Study-mean R2 (0.095) indicates limited generalizability. Mass concentration remains a practical reference within this evidence base; standardized reporting is needed, and geometry-derived metrics reflect administered, not delivered, dose.