Machine learning-driven prediction of eye irritation toxicity: Integration of in silico and in vitro study

Eye irritation (EI) toxicity poses critical challenges in chemical safety assessment, demanding alternatives to ethically contentious animal testing. We present the first integrative framework combining computational prediction with experimental validation for EI evaluation. Utilizing 5220 compounds characterized by 11 molecular fingerprints and 13 molecular descriptors (MDs), six machine learning (ML) algorithms generated 252 models through rigorous five-fold cross-validation. The optimized model demonstrated exceptional predictive accuracy (0.977 test-set performance), surpassing comparable approaches. Experimental validation assessed model-predicted compounds through in vitro corneal epithelial assays, with benchmarked biological responses confirming prediction reliability. This closed-loop methodology establishes a paradigm for ethical chemical safety assessment, achieving three critical advancements: replacement of animal testing through computational-experimental synergy, enhanced prediction reliability via machine learning-optimized feature selection, and accelerated developmental timelines through early-stage toxicity screening. The validated approach addresses interspecies extrapolation limitations while aligning with global 3Rs (Replacement, Reduction, Refinement) initiatives, offering transformative potential for next-generation risk assessment in chemical development.