Anastasia Tsyplakova1, Dr. Aleksandra Catic-Ðordevic2, Dr. Nikola Stefanovic2, Dr. Vangelis Karalis1
1Department of Pharmacy, School of Health Science, National and Kapodistrian University of Athens, 2Department of Pharmacy, Faculty of Medicine, University of Nis
Introduction: Mycophenolic acid (MPA) is a key immunosuppressant for renal transplant patients, typically used with calcineurin inhibitors and prednisone. It’s administered as mycophenolate mofetil (MMF), an ester prodrug of MPA that fully hydrolyzes to MPA, or as enteric-coated MPA. Its use requires careful individualization due to pharmacokinetic variability. Low MPA levels increase graft rejection risk, while high levels increase toxicity, such as gastrointestinal and hematological issues[1][2]. Population pharmacokinetics (PopPK) optimizes MPA therapy by identifying patient-specific factors for exposure-based dosing adjustments. However, traditional PopPK faces challenges like relying on simplistic models and struggling with individual variability, leading to difficulties in real-time predictions. Machine learning (ML) can improve predictive accuracy and reveal complex relationships between drug concentrations and clinical factors[3]. Integrating PopPK insights with ML will support a personalized MPA dosing strategy, reducing rejection and toxicity risks and enhancing transplant outcomes. Objective: This study aimed to: •Develop and validate two popPK models for MPA and MMF in renal transplant patients. •Apply ML techniques to identify and quantify predictive factors for MPA levels while uncovering potential correlations between MPA levels and clinical and biochemical parameters. •Examine the relationship between plasma and saliva MPA concentrations. Methods: The study included 80 renal transplant patients on MPA/MMF therapy, measuring MPA levels monthly in plasma and saliva. Each had one trough level per check-up after a steady state was achieved. Collected data included demographics, time post-transplantation (PTP), renal function (urea, creatinine), and haematological parameters (WBC, RBC, PLT). A PopPK analysis using nonlinear mixed-effects modelling in Monolix™ 2023R1 developed two models based on plasma MPA concentrations: a) Model 1, based on the MPA formulation, and b) Model 2, based on the MMF formulation. Model validation included both statistical and graphical methods. Intra-individual and inter-occasion (considering each checkup for the same individual as a distinct occasion) variability were accounted for in model selection. Inter-individual variability was analyzed through covariate effects on pharmacokinetic parameters. The ML techniques that were used referred to Principal Component Analysis (PCA) and ensemble methods (bagged and boosted Trees). PCA reduces dimensionality by converting correlated variables into uncorrelated principal components, maximizing variance. In this study, PCA was applied to visualize relationships among MPA dose, urea, PTP, MPA saliva levels, age, haematological parameters, and plasma MPA concentrations, identifying factors correlated with plasma MPA levels through loading vectors and clustering patterns. Ensemble learning improves prediction accuracy using multiple decision trees. Bagged trees train multiple decision trees on random subsamples of the dataset, reducing variance and enhancing stability. They can also estimate feature importance based on tree splits. Boosted trees iteratively build trees, correcting errors to improve accuracy and capturing nonlinear interactions[3]. The predictive performance of these models was evaluated using regression analysis. For each method, two models were created: one with saliva levels and one without. The ML methods codes were written in Python v.3.10.8 using the libraries “sklearn”, ” XGBoost”, ” CatBoost”, ” MLxtend”, “matplotlib”, and “seaborn”, among others. Results: Both PopPK models featured a one-compartment structure with first-order absorption and elimination kinetics. Total daily dose and PTP were significant covariates affecting clearance. The models showed low relative standard error and minimal inter-individual variability, confirming their robustness; goodness of fit plots highlighted their predictive ability. Machine learning analysis validated the findings and provided insights into individualized treatment. Among the ML methods, the best results were obtained from the PCA analysis and the application of ensemble method (boosting, bagging). Boosted and bagged trees analyses identified MPA (or MMF) dose as the main predictor, affirming its strong influence on MPA levels. Urea and PTP were also important predictors, with age showing moderate effect. Incorporating saliva concentration data improved predictive performance, with saliva emerging as the most influential variable, followed by dose, urea, and PTP.In contrast, WBC, PLT, and RBC made minimal contributions, suggesting a weaker link to MPA levels. Regression analysis yielded a coefficient of determination, R² value of more than 0.91, indicating strong predictive accuracy of the ensemble trees models. PCA confirmed these results, indicating dose as the primary determinant of MPA levels. Urea had a positive relationship with MPA concentration and an inverse relationship with PTP, WBC, RBC, and PLT, which clustered together, indicating a shared influence. Age was positioned between urea and PTP, showing a moderate but distinct impact. With saliva concentrations included, correlation patterns shifted slightly: MPA (plasma) remained closely tied to dose and urea, while saliva displayed a new but weaker correlation with MPA levels. The PTP, WBC, RBC, and PLT cluster stayed stable, reflecting a consistent influence, while urea and age showed similar relationships. Conclusion: This study integrated traditional popPK modeling with ML techniques to understand MPA/MMF pharmacokinetics better and enhance pharmacotherapy in transplant patients. It confirmed that MPA (or MMF) dose primarily influences MPA levels, as evidenced by bagged trees and PCA analyses, aligning with the pharmacokinetic principle of dose-dependent MPA exposure.[3] Urea and post-transplant time affected MPA variability, emphasizing the significance of patient-specific factors in immunosuppressive therapy[4]. The relationship of urea with MPA levels suggests altered metabolism or clearance in renal impairment cases[2]. Clustering of PTP, WBC, RBC, and PLT indicates hematological parameters reflect broader physiological changes post-transplant, but their effect on MPA levels is limited. Age moderately impacts drug levels, influenced by renal function and dosing[1]. The weak correlation between saliva and plasma concentrations indicates that saliva monitoring alone may be inadequate for MPA dosing adjustments, although saliva levels remain significant predictors of plasma levels, suggesting a complementary monitoring role. The developed models offer a practical method to personalize MPA (or MMF) dosing, ensuring accurate immunosuppressive treatment for kidney transplant patients[3][4].
[1] Greanya, E. D., Poulin, E., Partovi, N., Shapiro, R. J., Al-Khatib, M., & Ensom, M. H. H. (2012). Pharmacokinetics of tacrolimus and mycophenolate mofetil in renal transplant recipients on a corticosteroid-free regimen. American Journal of Health-System Pharmacy, 69(2), 134–142. https://doi.org/10.2146/ajhp110287 [2] González-Roncero, F., Gentil, M., Brunet, M., Algarra, G., Pereira, P., Cabello, V., & Peralvo, M. (2005). Pharmacokinetics of mycophenolate mofetil in kidney transplant patients with renal insufficiency. Transplantation Proceedings, 37(9), 3749–3751. https://doi.org/10.1016/j.transproceed.2005.09.202 [3] Woillard, J., Labriffe, M., Debord, J., & Marquet, P. (2021). Mycophenolic acid Exposure Prediction using Machine Learning. Clinical Pharmacology & Therapeutics, 110(2), 370–379. https://doi.org/10.1002/cpt.2216 [4] Wang, P., Xie, H., Zhang, Q., Tian, X., Feng, Y., Qin, Z., Yang, J., Shang, W., Feng, G., & Zhang, X. (2022). Population pharmacokinetics of mycophenolic acid in renal transplant patients: A comparison of the early and stable posttransplant stages. Frontiers in Pharmacology, 13. https://doi.org/10.3389/fphar.2022.859351
Reference: PAGE 33 (2025) Abstr 11711 [www.page-meeting.org/?abstract=11711]
Poster: Drug/Disease Modelling - Other Topics