2025 AIChE Annual Meeting

(207d) Enhancing Sensitivity Analysis of Complex Pharmaceutical Systems with Machine Learning

Authors

Roberto Andrea Abbiati, Boehringer Ingelheim Pharma & Co. KG
Davide Manca, Politecnico di Milano
Maria Papathanasiou, Imperial College London
Pharmacokinetic and pharmacodynamic (PK/PD) models play a crucial role in predicting drug transport within the body and assessing the efficacy and optimal dosage of treatments. These models incorporate kinetic parameters that govern transport across body compartments or influence drug efficacy, which can be linked to patient-specific characteristics.1 During parameter estimation model predictions are fitted to measurements from a heterogeneous population of patients. This generalisation can introduce sub-optimal decisions in the experimental design of pre-clinical trials or drug dosage formulation for subpopulations (i.e. pediatric or critically-ill patients), who have characteristics and drug-response behaviours that deviate from those of the average patient.2,3

Understanding how the parameter space influences model outputs is essential for correlating patient population variability with therapeutic outcome differences. Global Sensitivity Analysis (GSA) is a robust tool for examining parameter interactions, providing valuable insights into complex system dynamics. However, GSA performance is significantly impacted by model complexity, especially when dealing with large-scale and nonlinear PK/PD models that involve numerous parameters, leading to computational challenges. Strategies to simplify GSA, such as dividing parameters into subsets or using surrogate metamodels, often fall short as the parameter space expands.4,5

In this study, we explore the potential of Machine Learning (ML) to mitigate the complexity of PK/PD models by examining how hybridization levels influence GSA performance and evaluating whether surrogate utilization impacts model sensitivity to parametric uncertainty. Our results demonstrate that Feed-forward Neural Networks (NN) and Random Forest (RF) surrogates can accurately capture parameter interactions and sensitivities with minimal computational effort, requiring only a limited number of mechanistic model simulations. The NN-based surrogates exhibit high accuracy, reproducing sensitivity indices with a deviation of less than 3% compared to fully mechanistic models, while RF-based surrogates achieve similar results with slightly increased computational demands. The surrogate-enabled GSA achieves significant computational speed-up compared to traditional ODE-based approaches, enabling practical cohort-specific GSA for applications such as individualized treatment strategies and precision medicine. These results illustrate that the proposed surrogate-based GSA framework is particularly well-suited for large-scale and nonlinear models, including PBPK and QSP approaches, where it facilitates rapid parametric analyses and supports optimal study design for clinical and preclinical research.

References

(1) Abbiati, R. A.; Lamberti, G.; Grassi, M.; Trotta, F.; Manca, D. Definition and Validation of a Patient-Individualized Physiologically-Based Pharmacokinetic Model. Comput. Chem. Eng. 2016, 84, 394–408. https://doi.org/10.1016/j.compchemeng.2015.09.018.

(2) De Cock, R. F. W.; Piana, C.; Krekels, E. H. J.; Danhof, M.; Allegaert, K.; Knibbe, C. A. J. The Role of Population PK–PD Modelling in Paediatric Clinical Research. Eur. J. Clin. Pharmacol. 2011, 67 (1), 5–16. https://doi.org/10.1007/s00228-009-0782-9.

(3) Asín-Prieto, E.; Rodríguez-Gascón, A.; Isla, A. Applications of the Pharmacokinetic/Pharmacodynamic (PK/PD) Analysis of Antimicrobial Agents. J. Infect. Chemother. 2015, 21 (5), 319–329. https://doi.org/10.1016/j.jiac.2015.02.001.

(4) Sheikholeslami, R.; Razavi, S.; Gupta, H. V.; Becker, W.; Haghnegahdar, A. Global Sensitivity Analysis for High-Dimensional Problems: How to Objectively Group Factors and Measure Robustness and Convergence While Reducing Computational Cost. Environ. Model. Softw. 2019, 111, 282–299. https://doi.org/10.1016/j.envsoft.2018.09.002.

(5) Shan, S.; Wang, G. G. Metamodeling for High Dimensional Simulation-Based Design Problems. J. Mech. Des. 2010, 132 (051009). https://doi.org/10.1115/1.4001597.