XGBoost Machine Learning Model Accurately Predicts Early and Full Drug Release in PLGA Microspheres
Background
Developing long-acting injectable (LAI) formulations is crucial for improving patient adherence and therapeutic outcomes, especially for peptides. Poly(lactic-co-glycolic acid) (PLGA) microspheres are a preferred choice due to their biocompatibility and controlled hydrolytic degradation. However, optimizing these complex formulations typically demands extensive, time-consuming in vitro testing to characterize drug release kinetics. This iterative experimental process creates a significant bottleneck in drug development. A predictive model that can accurately forecast drug release profiles based on formulation parameters would drastically streamline the optimization of PLGA-based LAI systems.
Study Design
Researchers developed a two-stage machine learning framework using a published dataset comprising 321 release profiles from 89 drugs. The first stage involved a classification model designed to identify slow-release behavior (defined as ≤20% release within 3 days). The predicted early-release probability from this classifier was then integrated into a regression model to estimate the cumulative drug release over time. The framework was externally validated using in-house drug release data for olaparib-loaded PLGA microspheres and published data for semaglutide-based microspheres.
Results
Among various tree-based ensemble models, XGBoost demonstrated superior performance in predicting drug release from PLGA microspheres. It achieved the lowest mean absolute error (MAE = 0.126) and the highest Pearson correlation coefficient (r = 0.831), indicating strong predictive accuracy.
SHapley Additive exPlanations (SHAP)analysis revealed that drug and polymer molecular weight, the predictive slow-release probability, and polymer concentration were the most substantial factors influencing drug release behavior. External validation further confirmed the framework's robustness, yielding low MAE values of 0.096 for olaparib-loaded microspheres and 0.068 for semaglutide-based microspheres, underscoring its generalizability across different drug types.
Key Findings
- XGBoost model achieved MAE = 0.126 and r = 0.831 for predicting drug release.
- Drug molecular weight, polymer molecular weight, and polymer concentration are key influencers of release.
- External validation showed low MAE values: 0.096 for olaparib and 0.068 for semaglutide microspheres.
- The two-stage ML framework can predict both early-stage and full time-dependent drug release profiles.
Why It Matters
Accelerating the development and optimization of long-acting injectable (LAI) formulations is a critical practical takeaway for peptide users and formulators. This interpretable machine learning framework significantly reduces the need for extensive, time-consuming in vitro testing by accurately predicting drug release profiles. For novel peptide-loaded PLGA microspheres, like those for semaglutide or setmelanotide (as per domain context), this means potentially faster iteration cycles and more efficient material selection. The ability to predict release based on molecular and polymer properties could guide early-stage formulation design, leading to more cost-effective development of stable, controlled-release peptide therapeutics.
machine-learning
plga
drug-delivery
long-acting-injectable
semaglutide
olaparib