Machine Learning-Based PAT Tool for Real-Time Residual Moisture Monitoring in Lyophilized Biopharmaceuticals Using...

There is a growing demand in the pharmaceutical industry to optimize analytical methodologies, reducing time and resource consumption while ensuring rapid process development and cost-efficient manufacturing. Indeed, Food and Drug Administration (FDA) introduced the Process Analytical Technology (PAT) as an initiative enabling manufacturers to measure and control in real-time processes based on the Critical Quality Attributes (CQAs) of the product. This study explores the implementation of PAT as a strategic approach in the lyophilization process design and optimization. In this sense, Residual Moisture (RM) was selected as a CQA to evaluate lyophilized biopharmaceutical products, due to its impact on stability. Traditionally, RM quantification is obtained by Karl-Fischer (KF) titration, a destructive technique. Recently, Near-Infrared Spectroscopy (NIRS) was proposed as a non-destructive alternative; however, its effectiveness is hindered by spectral complexity and non-linear moisture-spectra relationships. Here, PAT tool integrating NIRS with Machine Learning (ML) is developed to enable rapid, at-line RM assessment, handling the complexity of non-linear relationships between RM values and spectral features. Different ML models - Partial Least Squares (PLS), Support Vector Regression (SVR), XGBoost, and an ensemble PLS-SVR approach - were trained on placebo samples and validated on drug products. SVR outperformed PLS at lower moisture levels, while PLS was more accurate at higher RM ranges. The ensemble model combining both approaches improved predictive accuracy across the full RM range tested. External validation using literature-sourced datasets, including placebo and simulated drug products with different excipients, confirmed the model's robustness and adaptability. Furthermore, a minimal dataset augmentation strategy enhanced model specialization for specific formulations. The proposed NIRS-ML ensemble model offers a scalable, non-destructive solution for real-time RM monitoring, supporting pharmaceutical manufacturing within the PAT framework. Its adaptability across formulations highlights its potential as a versatile tool for lyophilized biopharmaceutical manufacturing pipelines.