Cell Culture Media Release Using Inline Raman Spectroscopy and Artificial Neural Networks

Ensuring the quality and consistency of cell culture media is essential in biopharmaceutical manufacturing. This study investigates the application of inline Raman spectroscopy combined with machine learning algorithms for real-time characterization and release of cell culture media compositions. Raman spectroscopy, known for its ability to provide detailed molecular fingerprints through inelastic scattering, enables the noninvasive identification and quantification of Raman-active media components and the indirect estimation of certain non-Raman-active quality markers via correlation-based models. Our methodology involved the collection of Raman spectra from media mixtures with varying compositions, systematically altered through two experimental designs. These spectra were preprocessed and used to train Artificial Neural Networks (ANNs), which accurately predicted critical media markers based on both direct Raman signals and indirect correlations with Raman-detectable species, achieving R2 values of 0.988 (glucose), 0.985 (glutamine), 0.994 (osmolality), 0.994 (potassium), and 0.975 (sodium). Subsequently, K-Nearest Neighbors (KNN) models were employed to classify the media based on solution composition ranges. The KNN models achieved approximately 90% accuracy in classifying solution ranges, showcasing the potential of this combined approach for inline, real-time quality control of continous media preparations. This study underscores the effectiveness of integrating Raman spectroscopy and machine learning models within the Process Analytical Technology (PAT) framework to enhance media release and quality assurance in biopharmaceutical manufacturing.