Raman-based PAT for multi-attribute monitoring during VLP recovery by dual-stage CFF: attribute-specific spectral...

Spectroscopic soft sensors are developed by combining spectral data with chemometric modeling, and offer as Process Analytical Technology (PAT) tools powerful insights into biopharmaceutical processing. In this study, soft sensors based on Raman spectroscopy and linear or partial least squares (PLS) regression were developed and successfully transferred to a filtration-based recovery step of precipitated virus-like particles (VLPs). For near real-time monitoring of product accumulation and precipitant depletion, the dual-stage cross-flow filtration (CFF) set-up was equipped with an on-line loop in the second membrane stage. With this set-up, spectral data from three CFF runs were collected, differing in initial product concentration and process parameters. Under the scope of multi-attribute monitoring, a comprehensive investigation of the sensor sensitivity towards protein and precipitant and their Raman spectral features was carried out. This study reveals much higher sensitivity towards the precipitant ammonium sulfate (AMS) than the VLPs and the need for attribute-specific spectral preprocessing. To enhance the detector's sensitivity towards proteins, a higher exposure time was applied during CFF processing than during model building from pure-component stock solutions. As a result of this increased exposure time, the predominant sulfate band exhibited oversaturation effects, which otherwise could have been used for AMS quantification via linear regression. Nevertheless, AMS prediction using purpose-driven preprocessing operations and PLS models was achieved with normalization and a data-driven variable selection technique, next to baseline correction and signal smoothing. For VLP monitoring, a novel pre-cropping approach improved spectral appearance after further preprocessing in protein-associated wavenumber regions. However, fluctuations in prediction were much higher for VLPs than for AMS, and prediction accuracy was especially limited in low protein concentration ranges. These results highlight the potential of Raman-based PAT sensors for real-time monitoring of biopharmaceutical processes, while underscoring the general importance of attribute-specific selections of sensors, preprocessing operations, and models for PAT tool development.