Development of Raman Spectroscopy and Machine Learning Methods for Protein Aggregate Quantification: Application to...

Protein aggregation poses a significant risk to biopharmaceutical product quality, as even minor amounts of oligomeric species can compromise efficacy and safety. Rapid and reliable detection of protein aggregates thus remains a major challenge in biopharmaceutical manufacturing. Although traditional offline methods such as size-exclusion chromatography provide accurate results, their inherent time delays limit real-time process control capabilities. Consequently, there is an urgent scientific need for inline analytical techniques capable of selectively quantifying protein monomers and aggregates in real time to facilitate immediate corrective actions and enhance overall process robustness. Raman spectroscopy, as a tool for a process analytical technology application, is especially suitable due to its molecular specificity, rapid data acquisition, and compatibility with aqueous solutions commonly used in biopharmaceutical manufacturing. Addressing this need, this study establishes a Raman spectroscopy-based strategy for the selective detection and quantification of monomeric and aggregated forms of a model protein (bovine serum albumin). Controlled stress conditions were applied to generate aggregated species reproducibly, and a Latin Hypercube sampling design was used to independently vary protein concentration and aggregate fraction, ensuring that observed spectral effects were attributable to aggregation rather than concentration differences. Furthermore, spectral markers identified in spectra acquired from multiple chromatographic runs were qualitatively compared with offline reference measurements from size-exclusion chromatography. This limitation in real-time applicability was circumvented by chemometric machine learning approaches. The use of convolutional neural networks enabled the selective quantification of the protein monomers and aggregates and delivered superior predictive performance and robustness across cross-validation, independent testing, and synthetic perturbation scenarios compared to traditional chemometric approaches. Collectively, these results demonstrate that the selected Raman spectral markers, combined with advanced chemometric modeling, enable reliable, real-time monitoring of protein size variants in biopharmaceutical downstream processes.