Enhancing drug discovery with AI: Predictive modeling of pharmacokinetics using Graph Neural Networks and ensemble learning

Accurately predicting pharmacokinetic (PK) parameters such as absorption, distribution, metabolism, and excretion (ADME) is essential for optimizing drug efficacy, safety, and development timelines. Traditional experimental methods are often slow and expensive, driving the need for advanced AI-based approaches in PK modeling. This study compares cutting-edge machine learning models, including Graph Neural Networks (GNNs), Transformers, and Stacking Ensembles, against traditional models like Random Forest and XGBoost, using a dataset of over 10,000 bioactive compounds from the ChEMBL database. The Stacking Ensemble model achieved the highest accuracy (R² of 0.92, MAE of 0.062), outperforming GNNs (R² of 0.90) and Transformers (R² of 0.89). These AI models excelled in capturing complex molecular interactions and long-range dependencies, significantly improving PK predictions. The high accuracy achieved (R² ​= ​0.92) by the Stacking Ensemble method indicates that AI models can streamline the drug discovery process by reducing costly in vivo experiments, enabling faster go/no-go decisions during preclinical evaluations, and ultimately accelerating the development of new therapeutics. This reduction in time and cost could facilitate broader industry adoption of AI-driven PK modeling. Furthermore, Bayesian optimization was employed to fine-tune hyperparameters, further enhancing the performance and robustness of these predictive models.