An adaptive soft‐sensor for advanced real‐time monitoring of an antibody‐drug conjugation reaction

Abstract

In the production of antibody‐drug conjugates (ADCs), the conjugation reaction is a central step defining the final product composition and, hence, directly affecting product safety and efficacy. To enable real‐time monitoring, spectroscopic sensors in combination with multivariate regression models have gained popularity in recent years. The extended Kalman filter (EKF) can be used as so‐called soft‐sensor to fuse sensor predictions with long‐horizon forecasts by process models. This enables the dynamic update of the current state and provides increased robustness against experimental noise or model errors. Due to the uncertainty associated with sensor and process models in biopharmaceutical applications, the deployment of such soft‐sensors is challenging. In this study, we demonstrate the combination of an uncertainty‐aware sensor model with a kinetic reaction model using an EKF to monitor a site‐directed ADC conjugation reaction. As the sensor model, a Gaussian process regression model is presented to realize a time‐variant determination of the sensor uncertainty. The EKF fuses the time‐discrete predictions of the amount of conjugated drug from the sensor model with the time‐continuous predictions from the kinetic model. While the ADC species are not distinguishable by on‐line recorded UV/Vis spectra, the developed soft‐sensor is able to dynamically update all relevant reaction species. It could be shown that the use of time‐variant process and sensor noise computation approaches improved the performance of the EKF and achieved a reduction of the prediction error of up to 23% compared with the kinetic model. The developed framework proved to enhance robustness against noisy sensor measurements or wrong model initialization and was successfully transferred from batch to fed‐batch mode. In future, this framework could be implemented for model‐based process control and be adopted for other ADC conjugation reaction types.

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

“…By using a single-wavelength approach based on infinitesimal perturbations in Cui and Fearn (2018) , the studied CNNs were observed to have increased robustness compared to PLS models. A similar single-wavelength perturbation approach was used in Schiemer et al (2023) to perform automated variable selection for a Gaussian process regression model for monitoring an antibody-drug conjugation reaction.…”

“…Machine learning (ML) methods have gradually been applied to the field of chemometrics and have been shown to sometimes outperform linear methods on various regression and classification tasks, employing artificial neural networks (ANNs) ( Long et al, 1990 ; Santos et al, 2005 ), Gaussian process regression (GPR) ( Cui and Fearn, 2017 ; Malek et al, 2018 ), support vector machines (SVMs) ( Cui and Fearn, 2017 ), k-nearest neighbor (kNN) ( Wang et al, 2023 ) or convolutional neural networks (CNNs) ( Acquarelli et al, 2017 ; Bjerrum et al, 2017 ; Cui and Fearn, 2018 ; Blazhko et al, 2021 ; Passos and Mishra, 2021 ; Rolinger et al, 2021 ; Wang et al, 2023 ). Next to the increased accuracy, ML models were found to reduce the amount of preprocessing needed prior to spectral modeling ( Cui and Fearn, 2018 ; Rolinger et al, 2021 ; Tulsyan et al, 2021 ; Schiemer et al, 2023 ) and increase robustness against variability in the data ( Cui and Fearn, 2018 ; Yuanyuan and Zhibin, 2018 ). Major obstacles to successfully deploy these models for process monitoring in biopharmaceutical operations are the required amount of data for model calibration ( Tulsyan et al, 2019 ; Banner et al, 2021 ), the high number of hyperparameters ( Passos and Mishra, 2022 ) as well as the necessity for universally applicable diagnostic tools to reduce the black-box character of these models ( Burkart and Huber, 2021 ).…”

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

“…To effectively reduce the noise in the model, a larger data set preferably from fedbatch experiments would be required. Eventually, due to the nonlinear relationship of the Raman spectra and the VLP concentration, non-linear regression models should be evaluated such as kernelbased methods (Thissen et al, 2004;Barman et al, 2010;Zavala-Ortiz et al, 2020;Schiemer et al, 2023) or neural networks (Cui and Fearn, 2018;Wang et al, 2023;Schiemer et al, 2024).…”

Spectroscopic insights into multi-phase protein crystallization in complex lysate using Raman spectroscopy and a particle-free bypass

“…Due to high correlation within the data set, the spectra produced by the above stated techniques are commonly processed using chemometric techniques, e.g., principal component analysis (PCA) ( Simone et al, 2014a ), partial least squares (PLS) ( Simone et al, 2014a ) regression models, or gaussian process regression (GPR) ( Schiemer et al, 2023 ), just to name a few. Further explanations of chemometric methods can be found in published literature ( Wold et al, 2001 ; Chen et al, 2007 ; Bro and Smilde, 2014 ; Acquarelli et al, 2017 ).…”