Introduction: Blood–brain barrier (BBB) penetration is a crucial factor to assess in central nervous (CNS) system drug development. This criterion is commonly expressed as the unbound brain-to-plasma ratio (Kp,uu,brain) determined through in vivo and in vitro methods (e.g., microdialysis, brain slices). Nevertheless, these experimental tests are costly and time-consuming, limiting their application in early drug discovery.
Methods: In this study, we present a novel hierarchical support vector regression (HSVR) model to predict log Kp,uu,brain values directly from molecular descriptors. A curated dataset with experimentally measured Kp,uu,brain values was utilized to develop and validate the in silico models.
Results: The HSVR ensemble model achieved high predictive performance with no signs of overfitting across training, test and outlier sets. In addition, it demonstrated robust performance in mock testing with independent datasets, confirming its broad applicability. Importantly, HSVR outperformed commonly used models such as the Brain Penetration Predictor and Brain Exposure Efficiency Score. Beyond quantitative prediction, HSVR also accurately classified BBB+ versus BBB− compounds and achieved higher accuracy than SwissADME, lightBBB, and alvaRunner across multiple Kp,uu,brain cut-off thresholds.
Conclusions: Overall, HSVR provides a robust, interpretable, and cost-effective tool for predicting BBB penetration. It enables early decision-making, reduces reliance on animal testing, and accelerates the identification of CNS-targeted therapeutics.
