Sergio Sánchez Herrero1,2, Joaquin Herrerias-Lopez-De-Heredia2, Marina Cuquerella-Gilabert2, Jenifer Serna2, Almudena Rueda Ferreiro2, Laura Calvet3, Angel A. Juan4
1Dept. of Computer Science, Multimedia and Telecommunication, Universitat Oberta de Catalunya, 2Simulation Department, Empresarios Agrupados Internacional S.A., 3Telecommunications and Systems Engineering Department, Universitat Autònoma de Barcelona, 4Research Center on Production Management and Engineering, Universitat Politècnica de València
Introduction: In silico methods play a crucial role in drug development, enabling researchers to explore complex pharmacokinetic (PK) and pharmacodynamics (PD) behaviors without the need for extensive in vivo testing1. The increasing complexity of drug release profiles necessitates advanced modeling and simulation (M&S) tools that provide accurate, flexible, and scalable solutions2. PhysPK® is one such platform designed for precise drug release modeling. However, integrating machine learning (ML) techniques can significantly enhance the predictive power of these models3. Objectives: This study aims to implement diverse drug release profiles in PhysPK® and evaluate the efficacy of ML algorithms in optimizing dosage regimens. By combining mechanistic modeling with data-driven approaches, we seek to improve drug formulation strategies and enhance predictive accuracy. Methods: To capture a broad spectrum of drug release behaviors, we developed models incorporating Zero-order, First-order, Higuchi, Hixson-Crowell, Korsmeyer-Peppas, and Weibull kinetics4. Internal validation involved rigorous simulation-based analyses across theoretical scenarios to assess model reliability, while external validation utilized reference datasets to ensure real-world applicability. Various ML models—including Logistic Regression (LR), Artificial Neural Networks (ANN), Random Forest Regressor (RFR), Decision Tree Regressor, (DTR), LightGBM Regressor, XGBoost Regressor, K-Neighbors Regressor (KNN), and Support Vector Regression (SVR), were evaluated for their ability to refine dosage predictions and improve modified-release drug design. Key features analyzed include drug release parameters, PK profiles, and metrics such as AUC0-24, Tmax, Cmax, half-life (HL), volume of distribution (Vd), and clearance (Cl) Model performance was assessed using multiple metrics, including Average Fold Error (AFE), Absolute Average Fold Error (AAFE), Prediction Error (PE), Mean Squared Error (MSE), Mean Absolute Error (MAE), and R² score. Results: The drug release mathematical models, including Zero-order, First-order, Higuchi, Hixson–Crowell, Korsmeyer–Peppas, and Weibull, were successfully developed and validated in PhysPK®. Internal validation confirmed that these models accurately replicate drug release kinetics under various conditions, supporting their predictive reliability. External validation using literature data demonstrated high accuracy with minimal deviations from experimental values. AAFE values ranged between 0.8 and 1.25, indicating high predictive precision, while PPE remained below 2.5%, reflecting robust model reliability. ML integration further enhanced model performance, reducing optimization time and improving prediction accuracy for complex release mechanisms. Among the tested algorithms, XGBoost and LightGBM consistently outperformed others, achieving the lowest residual errors and highest R² values, underscoring their suitability for drug release modeling. The combination of mechanistic modeling with ML enabled more accurate parameter estimation and facilitated the identification of optimal drug formulations. Conclusion: PhysPK® proves to be a powerful tool for predicting drug release profiles and integrating these predictions into comprehensive PK/PD models. When coupled with ML techniques—particularly XGBoost and LightGBM—PhysPK® achieves superior predictive accuracy and robustness, significantly advancing computational pharmacokinetics. These findings establish a foundation for greater interchangeability between traditional pharmacokinetic modeling approaches and data-driven methodologies, paving the way for improved drug design and personalized medicine applications. The synergy between M&S frameworks and ML-driven optimization holds immense potential for streamlining drug development, reducing experimental costs, and accelerating regulatory approval processes.
[1] Gruber, Andrea, et al. “Prediction of human pharmacokinetics from chemical structure: combining mechanistic modeling with machine learning.” Journal of Pharmaceutical Sciences 113.1 (2024): 55-63. [2] Peppas, Nicholas A., and Balaji Narasimhan. “Mathematical models in drug delivery: How modeling has shaped the way we design new drug delivery systems.” Journal of Controlled Release 190 (2014): 75-81. [3] Sánchez-Herrero, Sergio, et al. “Embedding R inside the PhysPK Biosimulation Software for Pharmacokinetics Population Analysis.” BIO Integration 4.3 (2023). [4] Dash, Suvakanta, et al. “Kinetic modeling on drug release from controlled drug delivery systems.” Acta Pol Pharm 67.3 (2010): 217-223.
Reference: PAGE 33 (2025) Abstr 11660 [www.page-meeting.org/?abstract=11660]
Poster: Methodology – AI/Machine Learning