Evaluating Preclinical-to-Human Bioavailability Classification Using Smooth ROC Curve Estimation

Human oral bioavailability is a critical pharmacokinetic parameter that determines systemic drug exposure, informs dose selection, and guides compound prioritization during early-stage drug development. Although animal models are routinely employed to estimate human bioavailability, interspecies physiological differences and weak correlations often limit the accuracy of direct extrapolation. As an alternative, receiver operating characteristic (ROC) analysis provides a binary classification framework to evaluate the predictive performance of animal data. However, conventional empirical ROC curves are stepwise and discontinuous, which can complicate threshold determination-especially in small-sample settings. To enhance interpretability, this study applies a previously proposed smooth ROC estimation method based on Bernstein polynomials to a curated multispecies dataset that was initially analyzed using empirical techniques. Compared to the empirical approach, the Bernstein estimator produced smoother and more visually coherent ROC curves, though it slightly underestimated the area under the curve (AUC) for species with limited sample sizes. These findings reflect a trade-off between curve smoothness and discriminatory power, underscoring the practical utility of nonparametric smoothing techniques in translational pharmacokinetic modeling.