Application of near-infrared spectroscopy to bioreactor monitoring

Glucose, glutamine, lactate, and ammonia were found to be key analytes in cultures of Spodoptera frugiperda Sf-9 (Sf-9) and Trichoplusia ni BTI-Tn-5B1-4 (Tn-5B1-4) insect cells and PC-3 human prostate cancer cells. The respective uptake/production rates were determined for Sf-9 and Tn-5B1-4 insect cells during batch cultivations in a stirred tank bioreactor and for PC-3 human prostate cancer cells during batch cultivations in a perfusion rotating wall vessel (pRWV).; Near-infrared (NIR) spectroscopy, in combination with partial least-squares (PLS) regression analysis, was successfully applied to predict the key analytes in cell culture samples taken during cultivation of PC-3 human prostate cancer cells in tissue culture flasks. The respective PLS calibration models were built with single-beam spectra from randomly spiked tissue culture samples and analyte specific optimal spectral ranges within the 4800-4200 cm{dollar}sp{lcub}-1{rcub}{dollar} full spectral range. Random spiking has proven to be a valid methodology to remove cellular metabolism induced analyte correlation, thus enabling the use of actual culture samples in correlation free calibration model development. Analyte specific models based on optimized spectral ranges yielded standard errors of calibration (SEC) of 0.62 to 0.81 mM and standard errors of predictions (SEP) of 0.82 to 1.01 mM when applied to normal tissue culture flask samples. These errors would be acceptable for monitoring the analytes in the higher millimolar concentration range. The required number of PLS factors for these models was less than the corresponding models based on the full spectral range with respect to predictive ability. The developed spectral range optimization algorithm has proven to select spectral ranges containing characteristic analyte absorbance features.; Lactate levels are accurately measured on-line during cultivation of the cells in the pRWV with the model based on spiked tissue culture flask samples. Poor predictions resulted from the ammonia model. Digital Fourier filtering of the spectra prior to PLS regression improved all spectral monitoring results, although a prediction bias was still present for part of the glucose data and most of the glutamine and ammonia data. Accurate predictions of the concentration profiles demonstrate the analytical utility of NIR spectroscopy for bioreactor monitoring.