Tianqi Wang1, Tianyu Liu2, Somabha Mukherjee2, Matthias Wacker1
1National University of Singapore, Faculty of Science, Department of Pharmacy and Pharmaceutical Sciences, 2National University of Singapore, Faculty of Science, Department of Statistics & Data Science
Introduction Polymeric long-acting injectables (LAIs) offer significant benefits for the treatment of many chronic diseases, which extend therapeutic durations, reduce the dosing frequency, enhance patient compliance, and minimize the side effects associated with drug concentration fluctuations [1, 2]. Consequently, the evaluation of drug release kinetics—particularly burst release and release duration—is a critical quality attribute for LAIs [3]. However, traditional methods for assessing drug release are time-consuming and costly, highlighting the need for predictive models to accelerate formulation development [4, 5]. Machine learning models, for instance, facilitate outcome prediction by establishing relationships between input features and output results during model development. In the context of polymeric LAIs, where drug release profiles serve as the target output, input features such as drug/polymer properties, formulation characteristics, and release study conditions are carefully selected. A key challenge in this project lies in functional data prediction, which involves forecasting a series of unknown data points. This requires more complex models compared to single-outcome prediction. To address this, we integrated drug release models into the machine learning framework, leveraging their underlying mechanisms to simplify the predictive process. Additionally, the dataset of formulation release profiles is relatively small and poorly organized, with missing data further complicating database construction [4, 5]. To mitigate this, a clustering-based approach was employed for data imputation, ensuring both the quantity and quality of the data were maintained. This strategy enhances the robustness of the predictive model and supports more efficient formulation development. Objectives: -To compile and analyze a comprehensive dataset of in vitro release profiles for polymeric LAIs, while addressing missing data using advanced imputation techniques. -To develop and compare machine learning models for predicting drug release kinetics, evaluating their performance using robust metrics, and identifying key features influencing formulation release behavior. -To validate the predictive capability of the developed models through both retrospective and prospective studies. Methods A comprehensive systematic review was conducted to identify high-quality research articles reporting release studies for polymeric microspheres. Key features, including drug/polymer properties, formulation characteristics, and release study conditions, were identified and extracted to construct a robust database. During the data cleaning process, missing data were addressed using the k-nearest neighbors (KNN) algorithm for imputation. To model the complex, non-linear, and high-dimensional data, a decision tree learning approach and spectral clustering methods were employed. Specifically, the classification and regression tree (CART) algorithm, which utilizes Gini impurity for node splitting with categorical data, was implemented. Additionally, random forest and XGBoost models were trained to predict drug release kinetics. Model performance was evaluated using cross-validation, and further optimization was achieved through feature engineering and hyperparameter refinement. The predictive accuracy of the developed models was further assessed using both retrospective and prospective studies to ensure reliability and applicability. Results Over 1,500 in vitro release profiles of polymeric LAIs were compiled from published literature, alongside more than 30 numerical and categorical input features describing drug/polymer properties, formulation characteristics, and release study conditions. Approximately 10% of missing data were addressed using the KNN algorithm. To accommodate the hybrid output requirements and dataset size, three machine learning models, CART, random forest, and XGBoost, were tested and compared. Model performance was evaluated using multiple criteria, including the Akaike information criterion (AIC), coefficient of determination (R²), and mean absolute error (MAE). Validation through an outer (test) loop demonstrated strong agreement with the trained model, revealing that drug/polymer molecular weight, particle size, and drug load significantly influenced formulation release kinetics. Interestingly, release study settings, such as sampling volume and separation methods, also exhibited a notable impact on release behavior. The predictive capability of the model was further confirmed through prospective studies, utilizing data from both commercially available LAI formulations and those fabricated in-house. These findings underscore the model’s robustness and applicability in predicting drug release kinetics for polymeric LAIs. Discussion The in vitro release profile is essential for evaluating the performance and quality of LAIs, yet traditional methods are often inefficient. Our predictive model represents a promising alternative, potentially accelerating the formulation development and performance assessment. Challenges remain in accurately predicting functional data due to the lack of comprehensive databases and limited data accessibility. This study successfully leveraged the decision tree learning approach to elucidate relationships among drug and polymer features, formulation characteristics, and release conditions from extensive literature data. The model demonstrated robust predictive accuracy for newly formulated LAIs and provided insights into the underlying release mechanisms, deepening our understanding of LAI performance.
1. Abdelkader, H., Z. Fathalla, A. Seyfoddin, M. Farahani, T. Thrimawithana, A. Allahham, et al., Polymeric long-acting drug delivery systems (LADDS) for treatment of chronic diseases: Inserts, patches, wafers, and implants. Advanced Drug Delivery Reviews, 2021. 177: p. 113957. 2. Pandya, A.K., L.K. Vora, C. Umeyor, D. Surve, A. Patel, S. Biswas, et al., Polymeric in situ forming depots for long-acting drug delivery systems. Advanced Drug Delivery Reviews, 2023. 200: p. 115003. 3. Bao, Q., X. Wang, Y. Zou, Y. Wang, and D.J. Burgess, In vitro release testing method development for long-acting injectable suspensions. Int J Pharm, 2022. 622: p. 121840. 4. Bannigan, P., Z. Bao, R.J. Hickman, M. Aldeghi, F. Häse, A. Aspuru-Guzik, et al., Machine learning models to accelerate the design of polymeric long-acting injectables. Nature Communications, 2023. 14(1): p. 35. 5. Deng, J., Z. Ye, W. Zheng, J. Chen, H. Gao, Z. Wu, et al., Machine learning in accelerating microsphere formulation development. Drug Delivery and Translational Research, 2023. 13(4): p. 966-982.
Reference: PAGE 33 (2025) Abstr 11740 [www.page-meeting.org/?abstract=11740]
Poster: Methodology - New Modelling Approaches