(612e) UV-Vis Spectroscopy-Based PLS Inferential Sensor for Detection of Host Cell Proteins in Concentrated Antibody Solution

Continuous in-line process analytical technology (PAT) provides the information necessary for model based controllers such as model predictive control (MPC) to enable optimizing process operation and overcome process disturbances utilizing feedback from process measurements. Where direct measurements are unavailable, inferential sensors, also known as soft sensors, such as those based on partial least squares (PLS) models using ultraviolet-visible (UV-Vis) light spectroscopy can determine the concentration of individual protein species within a mixture as the spectral fingerprint of each protein is unique to its amino acid composition. While previous studies have demonstrated the advantages of utilizing inferential sensors at the monoclonal antibody (mAb) affinity chromatography step, their designed experiments were limited to simple mixtures of a few model proteins rather than complex mixtures of real host cell proteins (HCPs) and they did not explore concentration ranges relevant to the subsequent ion-exchange polishing step.

Here we report the performance of PLS models created to estimate the concentration of HCP impurities down to 1 mg/L in the presence of concentrated immunoglobulin G (IgG, up to 25 g/L) using UV-Vis spectroscopy measurements and a D-optimal experimental design. The development and implementation of the PLS model resulted in the concentration of HCP being estimated with a root mean squared error (RMSE) of 6.2 mg/L and the concentration of IgG estimated with an RMSE under 1 g/L. A wavelength-selection algorithm using PLS models with up to five components showed the most informative wavelengths for estimating the HCP concentration were in the range of 600 – 700 nm, 800 – 900 nm and 230 – 300 nm. Notably, by tracking the RMSE throughout, the PLS model accuracy towards HCP estimations was observed to improve, where an initial subset of approximately 40 wavelengths is initially added as model predictors, and then adding wavelengths beyond this subset was found to increase the RMSE. The greatest improvement in RMSE was seen upon adding a third wavelength where the RMSE decreased from approximately 12 mg/L to 8 mg/L. We show that while UV-Vis-based inferential sensors do not meet the performance of enzyme-linked immunosorbent assays (ELISA) or mass spectroscopy (MS) for estimating trace levels of HCP impurities, there is potential for inferential sensors being useful prior to the final polishing step to detect process disturbance and enable feedforward MPC systems.