Capacitance spectroscopy enables real-time monitoring of early cell death in mammalian cell culture

Adaptive modeling optimized by the data fusion strategy: Real-time dying cell percentage prediction using capacitance spectroscopy

The previous research showcased a partial least squares (PLS) regression model accurately predicting cell death percentages using in-line capacitance spectra. The current study advances the model accuracy through adaptive modeling employing a data fusion approach. This strategy enhances prediction performance by incorporating variables from the Cole-Cole model, conductivity and its derivatives over time, and Mahalanobis distance into the predictor matrix (X-matrix). Firstly, the Cole-Cole model, a mechanistic model with parameters linked to early cell death onset, was integrated to enhance prediction performance. Secondly, the inclusion of conductivity and its derivatives over time in the X-matrix mitigated prediction fluctuations resulting from abrupt conductivity changes during process operations. Thirdly, Mahalanobis distance, depicting spectral changes relative to a reference spectrum from a previous time point, improved model adaptability to independent test sets, thereby enhancing performance. The final data fusion model substantially decreased root-mean squared error of prediction (RMSEP) by around 50%, which is a significant boost in prediction accuracy compared to the prior PLS model. Robustness against reference spectrum selection was confirmed by consistent performance across various time points. In conclusion, this study illustrates that the data fusion strategy substantially enhances the model accuracy compared to the previous model relying solely on capacitance spectra.

An adaptive modeling approach using spiking-augmentation method to improve chemometric model performance in bioprocess monitoring

Intensified and continuous processes require fast and robust methods and technologies to monitor product titer for faster analytical turnaround time, process monitoring, and process control. The current titer measurements are mostly offline chromatography-based methods which may take hours or even days to get the results back from the analytical labs. Thus, offline methods will not meet the requirement of real time titer measurements for continuous production and capture processes. FTIR and chemometric based multivariate modeling are promising tools for real time titer monitoring in clarified bulk (CB) harvests and perfusate lines. However, empirical models are known to be vulnerable to unseen variability, specifically a FTIR chemometric titer model trained on a given biological molecule and process conditions often fails to provide accurate predictions of titer in another molecule under different process conditions. In this study, we developed an adaptive modeling strategy: the model was initially built using a calibration set of available perfusate and CB samples and then updated by augmenting spiking samples of the new molecules to the calibration set to make the model robust against perfusate or CB harvest of the new molecule. This strategy substantially improved the model performance and significantly reduced the modeling effort for new molecules.