In-line product quality monitoring during biopharmaceutical manufacturing using computational Raman spectroscopy

Raman-based PAT for VLP precipitation: systematic data diversification and preprocessing pipeline identification

Virus-like particles (VLPs) are a promising class of biopharmaceuticals for vaccines and targeted delivery. Starting from clarified lysate, VLPs are typically captured by selective precipitation. While VLP precipitation is induced by step-wise or continuous precipitant addition, current monitoring approaches do not support the direct product quantification, and analytical methods usually require various, time-consuming processing and sample preparation steps. Here, the application of Raman spectroscopy combined with chemometric methods may allow the simultaneous quantification of the precipitated VLPs and precipitant owing to its demonstrated advantages in analyzing crude, complex mixtures. In this study, we present a Raman spectroscopy-based Process Analytical Technology (PAT) tool developed on batch and fed-batch precipitation experiments of Hepatitis B core Antigen VLPs. We conducted small-scale precipitation experiments providing a diversified data set with varying precipitation dynamics and backgrounds induced by initial dilution or spiking of clarified Escherichia coli-derived lysates. For the Raman spectroscopy data, various preprocessing operations were systematically combined allowing the identification of a preprocessing pipeline, which proved to effectively eliminate initial lysate composition variations as well as most interferences attributed to precipitates and the precipitant present in solution. The calibrated partial least squares models seamlessly predicted the precipitant concentration with R 2 of 0.98 and 0.97 in batch and fed-batch experiments, respectively, and captured the observed precipitation trends with R 2 of 0.74 and 0.64. Although the resolution of fine differences between experiments was limited due to the observed non-linear relationship between spectral data and the VLP concentration, this study provides a foundation for employing Raman spectroscopy as a PAT sensor for monitoring VLP precipitation processes with the potential to extend its applicability to other phase-behavior dependent processes or molecules.

Simultaneous prediction of 16 quality attributes during protein A chromatography using machine learning based Raman spectroscopy models

Several key technologies for advancing biopharmaceutical manufacturing depend on the successful implementation of process analytical technologies that can monitor multiple product quality attributes in a continuous in-line setting. Raman spectroscopy is an emerging technology in the biopharma industry that promises to fit this strategic need, yet its application is not widespread due to limited success for predicting a meaningful number of quality attributes. In this study, we addressed this very problem by demonstrating new capabilities for preprocessing Raman spectra using a series of Butterworth filters. The resulting increase in the number of spectral features is paired with a machine learning algorithm and laboratory automation hardware to drive the automated collection and training of a calibration model that allows for the prediction of 16 different product quality attributes in an in-line mode. The demonstrated ability to generate these Raman-based models for in-process product quality monitoring is the breakthrough to increase process understanding by delivering product quality data in a continuous manner. The implementation of this multiattribute in-line technology will create new workflows within process development, characterization, validation, and control.

Realization of digital twin for dynamic control toward sample variation of ion exchange chromatography in antibody separation

Digital twin (DT) is a virtual and digital representation of physical objects or processes. In this paper, this concept is applied to dynamic control of the collection window in the ion exchange chromatography (IEC) toward sample variations. A possible structure of a feedforward model-based control DT system was proposed. Initially, a precise IEC mechanistic model was established through experiments, model fitting, and validation. The average root mean square error (RMSE) of fitting and validation was 8.1% and 7.4%, respectively. Then a model-based gradient optimization was performed, resulting in a 70.0% yield with a remarkable 11.2% increase. Subsequently, the DT was established by systematically integrating the model, chromatography system, online high-performance liquid chromatography, and a server computer. The DT was validated under varying load conditions. The results demonstrated that the DT could offer an accurate control with acidic variants proportion and yield difference of less than 2% compared to the offline analysis. The embedding mechanistic model also showed a positive predictive performance with an average RMSE of 11.7% during the DT test under >10% sample variation. Practical scenario tests indicated that tightening the control target could further enhance the DT robustness, achieving over 98% success rate with an average yield of 72.7%. The results demonstrated that the constructed DT could accurately mimic real-world situations and perform an automated and flexible pooling in IEC. Additionally, a detailed methodology for applying DT was summarized.

Application of Raman spectroscopy during pharmaceutical process development for determination of critical quality attributes in Protein A chromatography

Raman spectroscopy is considered a Process Analytical Technology (PAT) tool in biopharmaceutical downstream processes. In the past decade, researchers have shown Raman spectroscopy's feasibility in determining Critical Quality Attributes (CQAs) in bioprocessing. This study verifies the feasibility of implementing a Raman-based PAT tool in Protein A chromatography as a CQA monitoring technique, for the purpose of accelerating process development and achieving real-time release in manufacturing. A system connecting Raman to a Tecan liquid handling station enables high-throughput model calibration. One calibration experiment collects Raman spectra of 183 samples with 8 CQAs within 25 h. After applying Butterworth high-pass filters and k-nearest neighbor (KNN) regression for model training, the model showed high predictive accuracy for fragments (Q2 = 0.965) and strong predictability for target protein concentration, aggregates, as well as charge variants (Q2≥ 0.922). The model's robustness was confirmed by varying the elution pH, load density, and residence time using 19 external validation preparative Protein A chromatography runs. The model can deliver elution profiles of multiple CQAs within a set point ± 0.3 pH range. The CQA readouts were presented as continuous chromatograms with a resolution of every 28 s for enhanced process understanding. In external validation datasets, the model maintained strong predictability especially for target protein concentration (Q2 = 0.956) and basic charge variants (Q2 = 0.943), except for overpredicted HCP (Q2 = 0.539). This study demonstrates a rapid, effective method for implementing Raman spectroscopy for in-line CQA monitoring in process development and biomanufacturing, eliminating the need for labor-intensive sample pooling and handling.

Spectral insights: Navigating the frontiers of biomedical and microbiological exploration with Raman spectroscopy

Raman spectroscopy (RS), a powerful analytical technique, has gained increasing recognition and utility in the fields of biomedical and biological research. Raman spectroscopic analyses find extensive application in the field of medicine and are employed for intricate research endeavors and diagnostic purposes. Consequently, it enjoys broad utilization within the realm of biological research, facilitating the identification of cellular classifications, metabolite profiling within the cellular milieu, and the assessment of pigment constituents within microalgae. This article also explores the multifaceted role of RS in these domains, highlighting its distinct advantages, acknowledging its limitations, and proposing strategies for enhancement.

Generative data augmentation and automated optimization of convolutional neural networks for process monitoring

Chemometric modeling for spectral data is considered a key technology in biopharmaceutical processing to realize real-time process control and release testing. Machine learning (ML) models have been shown to increase the accuracy of various spectral regression and classification tasks, remove challenging preprocessing steps for spectral data, and promise to improve the transferability of models when compared to commonly applied, linear methods. The training and optimization of ML models require large data sets which are not available in the context of biopharmaceutical processing. Generative methods to extend data sets with realistic in silico samples, so-called data augmentation, may provide the means to alleviate this challenge. In this study, we develop and implement a novel data augmentation method for generating in silico spectral data based on local estimation of pure component profiles for training convolutional neural network (CNN) models using four data sets. We simultaneously tune hyperparameters associated with data augmentation and the neural network architecture using Bayesian optimization. Finally, we compare the optimized CNN models with partial least-squares regression models (PLS) in terms of accuracy, robustness, and interpretability. The proposed data augmentation method is shown to produce highly realistic spectral data by adapting the estimates of the pure component profiles to the sampled concentration regimes. Augmenting CNNs with the in silico spectral data is shown to improve the prediction accuracy for the quantification of monoclonal antibody (mAb) size variants by up to 50% in comparison to single-response PLS models. Bayesian structure optimization suggests that multiple convolutional blocks are beneficial for model accuracy and enable transfer across different data sets. Model-agnostic feature importance methods and synthetic noise perturbation are used to directly compare the optimized CNNs with PLS models. This enables the identification of wavelength regions critical for model performance and suggests increased robustness against Gaussian white noise and wavelength shifts of the CNNs compared to the PLS models.