Integration of just-in-time learning with variational autoencoder for cell culture process monitoring based on Raman spectroscopy.

Protein production in the biopharmaceutical industry necessitates the utilization of multiple analytical techniques and control methodologies to ensure both safety and consistency. To facilitate real-time monitoring and control of cell culture processes, Raman spectroscopy has emerged as a versatile analytical technology. This technique, categorized as a Process Analytical Technology, employs chemometric models to establish correlations between Raman signals and key variables of interest. One notable approach for achieving real-time monitoring is through the application of just-in-time learning (JITL), an industrial soft sensor modeling technique that utilizes Raman signals to estimate process variables promptly. The conventional Raman-based JITL method relies on the K-nearest neighbor (KNN) algorithm with Euclidean distance as the similarity measure. However, it falls short of addressing the impact of data uncertainties. To rectify this limitation, this study endeavors to integrate JITL with a variational autoencoder (VAE). This integration aims to extract dominant Raman features in a nonlinear fashion, which are expressed as multivariate Gaussian distributions. Three experimental runs using different cell lines were chosen to compare the performance of the proposed algorithm with commonly utilized methods in the literature. The findings indicate that the VAE-JITL approach consistently outperforms partial least squares, convolutional neural network, and JITL with KNN similarity measure in accurately predicting key process variables.

Cell culture product quality attribute prediction using convolutional neural networks and Raman spectroscopy.

Advanced process control in the biopharmaceutical industry often lacks real-time measurements due to resource constraints. Raman spectroscopy and Partial Least Squares (PLS) models are often used to monitor bioprocess cultures in real-time. In spite of the ease of training, the accuracy of the PLS model is impacted if it is not used to predict quality attributes for the cell lines it is trained on. To address this issue, a deep convolutional neural network (CNN) is proposed for offline modeling of metabolites using Raman spectroscopy. By utilizing asymmetric least squares smoothing to adjust Raman spectra baselines, a generic training data set is created by amalgamating spectra from various cell lines and operating conditions. This data set, combined with their derivatives, forms a two-dimensional model input. The CNN model is developed and validated for predicting different quality variables against measurements from various continuous and fed-batch experimental runs. Validation results confirm that the deep CNN model is an accurate generic model of the process to predict real-time quality attributes, even in experimental runs not included in the training data. This model is robust and versatile, requiring no recalibration when deployed at different sites to monitor various cell lines and experimental runs.

Machine learning-based model predictive controller design for cell culture processes.

The biopharmaceutical industry continuously seeks to optimize the critical quality attributes to maintain the reliability and cost-effectiveness of its products. Such optimization demands a scalable and optimal control strategy to meet the process constraints and objectives. This work uses a model predictive controller (MPC) to compute an optimal feeding strategy leading to maximized cell growth and metabolite production in fed-batch cell culture processes. The lack of high-fidelity physics-based models and the high complexity of cell culture processes motivated us to use machine learning algorithms in the forecast model to aid our development. We took advantage of linear regression, the Gaussian process and neural network models in the MPC design to maximize the daily protein production for each batch. The control scheme of the cell culture process solves an optimization problem while maintaining all metabolites and cell culture process variables within the specification. The linear and nonlinear models are developed based on real cell culture process data, and the performance of the designed controllers is evaluated by running several real-time experiments.

Automatic real-time calibration, assessment, and maintenance of generic Raman models for online monitoring of cell culture processes.

Raman spectroscopy is a multipurpose analytical technology that has found great utility in real-time monitoring and control of critical performance parameters of cell culture processes. As a process analytical technology (PAT) tool, the performance of Raman spectroscopy relies on chemometric models that correlate Raman signals to the parameters of interest. The current calibration techniques yield highly specific models that are reliable only on the operating conditions they are calibrated in. Furthermore, once models are calibrated, it is typical for the model performance to degrade over time due to various recipe changes, raw material variability, and process drifts. Maintaining the performance of industrial Raman models is further complicated due to the lack of a systematic approach to assessing the performance of Raman models. In this article, we propose a real-time just-in-time learning (RT-JITL) framework for automatic calibration, assessment, and maintenance of industrial Raman models. Unlike traditional models, RT-JITL calibrates generic models that can be reliably deployed in cell culture experiments involving different modalities, cell lines, media compositions, and operating conditions. RT-JITL is a first fully integrated and fully autonomous platform offering a self-learning approach for calibrating and maintaining industrial Raman models. The efficacy of RT-JITL is demonstrated on experimental studies involving real-time predictions of various cell culture performance parameters, such as metabolite concentrations, viability, and viable cell density. RT-JITL framework introduces a paradigm shift in the way industrial Raman models are calibrated, assessed, and maintained, which to the best of authors' knowledge, have not been done before.

A machine-learning approach to calibrate generic Raman models for real-time monitoring of cell culture processes.

The manufacture of biotherapeutic proteins consists of complex upstream unit operations requiring multiple raw materials, analytical techniques, and control strategies to produce safe and consistent products for patients. Raman spectroscopy is a ubiquitous multipurpose analytical technique in biopharmaceutical manufacturing for real-time predictions of critical parameters in cell culture processes. The accuracy of Raman spectroscopy relies on chemometric models that need to be carefully calibrated. The existing calibration procedure is nontrivial to implement as it necessitates executing multiple carefully designed experiments for generating relevant calibration sets. Further, existing procedure yields calibration models that are reliable only in operating conditions they were calibrated in. This creates a unique challenge in clinical manufacturing where products have limited production history. In this paper, a novel machine-learning procedure based on just-in-time learning (JITL) is proposed to calibrate Raman models. Unlike traditional techniques, JITL-based generic Raman models can be reliably used for different modalities, cell lines, culture media, and operating conditions. The accuracy of JITL-based generic models is demonstrated on several validation studies involving real-time predictions of critical cell culture performance parameters, such as glucose, glutamate, glutamine, ammonium, lactate, sodium, calcium, viability, and viable cell density. The proposed JITL framework introduces a paradigm shift in the way industrial Raman models are calibrated, which to the best of authors' knowledge have not been done before.

Advances in industrial biopharmaceutical batch process monitoring: Machine-learning methods for small data problems.

Biopharmaceutical manufacturing comprises of multiple distinct processing steps that require effective and efficient monitoring of many variables simultaneously in real-time. The state-of-the-art real-time multivariate statistical batch process monitoring (BPM) platforms have been in use in recent years to ensure comprehensive monitoring is in place as a complementary tool for continued process verification to detect weak signals. This article addresses a longstanding, industry-wide problem in BPM, referred to as the "Low-N" problem, wherein a product has a limited production history. The current best industrial practice to address the Low-N problem is to switch from a multivariate to a univariate BPM, until sufficient product history is available to build and deploy a multivariate BPM platform. Every batch run without a robust multivariate BPM platform poses risk of not detecting potential weak signals developing in the process that might have an impact on process and product performance. In this article, we propose an approach to solve the Low-N problem by generating an arbitrarily large number of in silico batches through a combination of hardware exploitation and machine-learning methods. To the best of authors' knowledge, this is the first article to provide a solution to the Low-N problem in biopharmaceutical manufacturing using machine-learning methods. Several industrial case studies from bulk drug substance manufacturing are presented to demonstrate the efficacy of the proposed approach for BPM under various Low-N scenarios.

Assessment of type II diabetes mellitus using irregularly sampled measurements with missing data.

Diabetes mellitus is one of the leading diseases in the developed world. In order to better regulate blood glucose in a diabetic patient, improved modelling of insulin-glucose dynamics is a key factor in the treatment of diabetes mellitus. In the current work, the insulin-glucose dynamics in type II diabetes mellitus can be modelled by using a stochastic nonlinear state-space model. Estimating the parameters of such a model is difficult as only a few blood glucose and insulin measurements per day are available in a non-clinical setting. Therefore, developing a predictive model of the blood glucose of a person with type II diabetes mellitus is important when the glucose and insulin concentrations are only available at irregular intervals. To overcome these difficulties, we resort to online sequential Monte Carlo (SMC) estimation of states and parameters of the state-space model for type II diabetic patients under various levels of randomly missing clinical data. Our results show that this method is efficient in monitoring and estimating the dynamics of the peripheral glucose, insulin and incretins concentration when 10, 25 and 50% of the simulated clinical data were randomly removed.