Half-life prediction of central nervous system (CNS) small molecules in humans using gradient tree boosting.

Abstract

Aims: To develop a machine learning (ML) model for early-stage prediction of human half-life of oral central nervous system (CNS) drugs and to establish a curated dataset, including key in vitro and in vivo data, to support future modeling efforts.

Materials & methods: Human and rat half-life, plasma protein binding (PPB), and liver microsomal clearance (LM) data for 76 diverse CNS drugs and candidates were obtained from public sources or evaluated at WuXi AppTec. Gradient tree boosting (GTB) models were constructed using ChemAxon's Trainer Engine. Feature importance was assessed, and model performance was evaluated on an external validation set.

Results: The best-performing model achieved 82.4% of predictions within two-fold of observed values, with a coefficient of determination (R²) of 0.75 and a root mean square error (RMSE) of 0.25. Good generalizability was confirmed using similarity-based data splitting and Y-randomization. Integration of in vitro features, preclinical in vivo data, and physicochemical properties substantially improved predictive performance. Key features driving accurate human half-life prediction were identified.

Conclusion: This model demonstrates practical applications for early-stage prediction of human half-life and prioritization of CNS drug candidates. The curated dataset offers a valuable resource to enhance internal databases and advance more robust predictive models.