Species-specific q-RASAR models for PFAS rodent oral toxicity: Interpretable prediction and large-scale data-gap filling

Per- and polyfluoroalkyl substances (PFAS) comprise thousands of persistent chemicals for which mammalian oral toxicity data remain sparse, limiting risk assessment and prioritization. Best Subset Selection (BSS) was used to identify optimal descriptor sets for baseline quantitative structure-activity relationship (QSAR) models; read-across-derived similarity and error descriptors were then integrated to construct quantitative read-across structure-activity relationship (q‑RASAR) models with improved predictivity. Experimental LD 50 values (mg/kg) were standardized and transformed to pLD 50 (mmol/kg) to harmonize endpoints, followed by 1D/2D descriptor generation (PaDEL) and genetic algorithm-driven feature selection. For rats, the final q-RASAR combined three structural descriptors with a similarity concordance metric and achieved strong goodness-of-fit and predictivity ( R 2 = 0.92; Q 2 LOO = 0.90; Q 2 F1 = 0.90). For mice, the final q-RASAR relied on four kernel similarity descriptors with comparable predictive performance ( R 2 = 0.93; Q 2 LOO = 0.90; Q 2 F1 = 0.89). Applicability domain assessment and Y-randomization confirmed model robustness and non-random correlations. Mechanistically, the rat model indicates that aromatic amine fragments and hydrogen-bond interaction capacity, together with similarity concordance, are key drivers of increased toxicity, whereas a mass-distribution autocorrelation descriptor contributes complementary structural variance. In the mouse model, toxicity is governed primarily by chemical neighborhood behavior, where similarity to active analogs elevates predicted toxicity and similarity to inactive analogs reduces it. The best-performing q-RASAR models were applied to 2500+ PFAS from the NORMAN SusDat database to support regulatory-style screening and triage. External predictions suggested that more toxic candidates were enriched in functionalized or heteroatom-containing PFAS and certain long-chain/activated derivatives, whereas less toxic candidates were dominated by simpler aliphatic fluorinated ethers/esters and related perfluoroalkyl architectures.