The introduction of the QbD and PAT guidelines by the FDA have been a strong driving force in industry and academia to improve process understanding and to implement real-time monitoring of critical quality attributes in biopharmaceutical production processes. Still, process monitoring and decisions in downstream processing are generally based on univariate signals. The information content of a univariate signal is however limited and important deviations may only be detected by offline analytics. To overcome these limitations, the evaluation of spectral measurements in combination with multivariate data analysis has been introduced previously as a powerful tool with many possible applications. Especially UV/Vis absorption spectra are promising for quantifying co-eluting proteins in real-time. Through partial least-squares (PLS) regression models, multicomponent mixtures can be deconvoluted to single protein concentrations. This project illustrates data for the real-time deconvolution of chromatograms for industrial applications. With a custom-made software toolchain, decisions are then taken to control effluent pooling. However, the performance of PLS models can drop drastically if unexpected species occur. Additionally, the model calibration can be tedious and the resulting PLS model may be system and detector dependent. To address these drawbacks, an alternative method was investigated making use of Lambert-Beer’s law. The method does not require offline analytics but deconvolutes absorbance data autonomously into concentrations and absorption spectra. With multivariate curve resolution (MCR), chromatograms and spectra were retrieved for multiple co-eluting species. The retrieved spectra can provide valuable information to identify components. The results were compared to results from offline fraction analytics. Chromatograms for two and three component mixtures were deconvoluted and analyzed. Under process conditions, PLS models are well suited for monitoring due to their higher accuracies. MCR may however reduce the dependencies on offline analytics during process development, and circumvent the analytical bottleneck to a certain degree. It may further be useful for detecting and monitoring out-of-specification components during both process development and monitoring. In summary, classical PLS regression and the newly proposed method are complementary and may each be applied in different circumstances.

Spectral deconvolution of chromatograms with PLS models and calibration-free methods

251st National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.