Dissecting corneal drug permeability under experimental variability and drug cold-start challenges via hybrid machine...

Corneal drug permeability is a key determinant of topical ocular drug delivery efficiency, yet its quantitative prediction remains challenging due to experimental noise and limited generalizability across drugs. Herein, we systematically integrated ex vivo drug and formulation corneal permeability data and developed a hybrid machine learning framework that jointly incorporates intrinsic and formulation-level information to quantitatively characterize corneal drug permeation under high-noise and drug cold-start conditions. We curated the first comprehensive ex vivo corneal apparent permeability coefficient (P_app) dataset containing 575 formulation records covering 65 active pharmaceutical ingredients (APIs), 85 penetration enhancers, and 4 animal species. We also curated a molecular dataset of 205 drugs representing intrinsic permeation potential, which was used as prior information for formulation-level prediction. Systematic comparison across multiple machine learning algorithms showed that Tabular Prior-data Fitted Network (TabPFN) achieved the best performance, with R² values of 0.54 for molecular permeability and 0.88 for formulation-level prediction. However, performance degraded in scenarios involving unseen drugs. Incorporating molecular permeability as a prior feature reduced cold-start error by 14.6% and achieved an MAE of 0.42 for logP_app, approaching the empirical upper bound of experimental reproducibility (MAE = 0.33) as estimated from cross-source measurements. These results demonstrate that molecular-level information substantially improves model generalization under drug cold-start scenarios. Interpretability analyses indicate that intrinsic molecular permeation establishes a baseline permeability, while formulation components introduce nonlinear, mechanism-specific modulation. Simulation results further demonstrated efficient few-shot generalization across unseen drugs and rapidly improved predictive performance using only limited additional experimental data. Overall, this study establishes an interpretable, hybrid corneal permeability modeling framework for the rational design of ophthalmic formulations. By offering a generalizable paradigm for predicting formulation performance under cross-drug conditions, this framework is poised to benefit the field by enhancing animal welfare and improving drug development efficiency.