An innovative hybrid modeling approach for simultaneous prediction of cell culture process dynamics and product quality.

The use of hybrid models is extensively described in the literature to predict the process evolution in cell cultures. These models combine mechanistic and machine learning methods, allowing the prediction of complex process behavior, in the presence of many process variables, without the need to collect a large amount of data. Hybrid models cannot be directly used to predict final product critical quality attributes, or CQAs, because they are usually measured only at the end of the process, and more mechanistic knowledge is needed for many classes of CQAs. The historical models can instead predict the CQAs better; however, they cannot directly relate manipulated process parameters to final CQAs, as they require knowledge of the process evolution. In this work, we propose an innovative modeling approach based on combining a hybrid propagation model with a historical data-driven model, that is, the combined hybrid model, for simultaneous prediction of full process dynamics and CQAs. The performance of the combined hybrid model was evaluated on an industrial dataset and compared to classical black-box models, which directly relate manipulated process parameters to CQAs. The proposed combined hybrid model outperforms the black-box model by 33% on average in predicting the CQAs while requiring only around half of the data for model training to match performance. Thus, in terms of model accuracy and experimental costs, the combined hybrid model in this study provides a promising platform for process optimization applications.

Design of Biopharmaceutical Formulations Accelerated by Machine Learning.

In addition to activity, successful biological drugs must exhibit a series of suitable developability properties, which depend on both protein sequence and buffer composition. In the context of this high-dimensional optimization problem, advanced algorithms from the domain of machine learning are highly beneficial in complementing analytical screening and rational design. Here, we propose a Bayesian optimization algorithm to accelerate the design of biopharmaceutical formulations. We demonstrate the power of this approach by identifying the formulation that optimizes the thermal stability of three tandem single-chain Fv variants within 25 experiments, a number which is less than one-third of the experiments that would be required by a classical DoE method and several orders of magnitude smaller compared to detailed experimental analysis of full combinatorial space. We further show the advantage of this method over conventional approaches to efficiently transfer historical information as prior knowledge for the development of new biologics or when new buffer agents are available. Moreover, we highlight the benefit of our technique in engineering multiple biophysical properties by simultaneously optimizing both thermal and interface stabilities. This optimization minimizes the amount of surfactant in the formulation, which is important to decrease the risks associated with corresponding degradation processes. Overall, this method can provide high speed of converging to optimal conditions, the ability to transfer prior knowledge, and the identification of new nonlinear combinations of excipients. We envision that these features can lead to a considerable acceleration in formulation design and to parallelization of operations during drug development.

Knowledge transfer across cell lines using hybrid Gaussian process models with entity embedding vectors.

To date, a large number of experiments are performed to develop a biochemical process. The generated data is used only once, to take decisions for development. Could we exploit data of already developed processes to make predictions for a novel process, we could significantly reduce the number of experiments needed. Processes for different products exhibit differences in behaviour, typically only a subset behave similar. Therefore, effective learning on multiple product spanning process data requires a sensible representation of the product identity. We propose to represent the product identity (a categorical feature) by embedding vectors that serve as input to a Gaussian process regression model. We demonstrate how the embedding vectors can be learned from process data and show that they capture an interpretable notion of product similarity. The improvement in performance is compared to traditional one-hot encoding on a simulated cross product learning task. All in all, the proposed method could render possible significant reductions in wet-lab experiments.

Hybrid Models for the simulation and prediction of chromatographic processes for protein capture.

The biopharmaceutical industries are continuously faced with the pressure to reduce the development costs and accelerate development time scales. The traditional approach of heuristic-based or platform process-based optimization is soon getting obsolete, and more generalized tools for process development and optimization are required to keep pace with the emerging trends. Thus, advanced model-based methods that can reduce the can ensure accelerated development of robust processes with minimal experiments are necessary. Though mechanistic models for chromatography are quite popular, their success is limited by the need to have accurate knowledge of adsorption isotherms and mass transfer kinetics. As an alternative, in this work, a hybrid modeling approach is proposed. Thereby, the chromatographic unit behavior is learned by a combination of neural network and mechanistic model while fitting suitable experimental breakthrough curves. Since this approach does not require identifying suitable mechanistic assumptions for all the phenomena, it can be developed with lower effort. Thus, allowing the scientists to concentrate their focus on process development. The performance of the hybrid model is compared with the mechanistic Lumped kinetic Model for in-silico data and experiments conducted on a system of industrial relevance. The flexibility of the hybrid modeling approach results in about three times higher accuracies compared to Lumped Kinetic Model. This is validated for five different isotherm models used to simulate data, with the hybrid model showing about two to three times lower prediction errors in all the cases. Not only in prediction, but we could also show that the hybrid model is more robust in extrapolating across process conditions with about three times lower error than the LKM. Additionally, it could be demonstrated that an appropriately tailored formulation of the hybrid model can be used to generate representations for the underlying principles such as adsorption equilibria and mass transfer kinetics.

Machine Learning for Biologics: Opportunities for Protein Engineering, Developability, and Formulation.

Successful biologics must satisfy multiple properties including activity and particular physicochemical features that are globally defined as developability. These multiple properties must be simultaneously optimized in a very broad design space of protein sequences and buffer compositions. In this context, artificial intelligence (AI), and especially machine learning (ML), have great potential to accelerate and improve the optimization of protein properties, increasing their activity and safety as well as decreasing their development time and manufacturing costs. We highlight the emerging applications of ML in biologics discovery and development, focusing on protein engineering, early biophysical screening, and formulation. We discuss the power of ML in extracting information from complex datasets and in reducing the necessary experimental effort to simultaneously achieve multiple quality targets. We finally anticipate possible future interventions of AI in several steps of the biological landscape.

Model based strategies towards protein A resin lifetime optimization and supervision.

The high cost of protein A resins drives the biopharmaceutical industry to maximize its lifetime, which is limited by several processes, usually referred to as resin aging. In this work, two model based strategies are presented, aiming to control and improve the resin lifetime. The first approach, purely statistical, enables qualitative monitoring of the column state and prediction of column performance indicators (e.g. yield, purity and dynamic binding capacity) from chromatographic on-line data (e.g. UV signal). The second one, referred to as hybrid modeling, is based on a lumped kinetic model, which includes two aging parameters fitted on several resin cycling experimental campaigns with varying cleaning procedures (CP). The first aging parameter accounts for binding capacity deterioration (caused by ligand degradation, leaching, and pore occlusion), while the second accounts for a decreased mass transfer rate (mainly caused by fouling). The hybrid model provides important insights into the prevailing aging mechanism as a function of the different CPs. In addition, it can be applied to model based CP optimization and yield forecasting with the capability of state estimation corrections based on on-line process information. Both approaches show promising results, which could help to significantly extend the resin lifetime. This comes along with increased understanding, reduced experimental effort, decreased cost of goods, and improved process robustness.

Cell culture process metabolomics together with multivariate data analysis tools opens new routes for bioprocess development and glycosylation prediction.

Multivariate latent variable methods have become a popular and versatile toolset to analyze bioprocess data in industry and academia. This work spans such applications from the evaluation of the role of the standard process variables and metabolites to the metabolomics level, that is, to the extensive number metabolic compounds detectable in the extracellular and intracellular domains. Given the substantial effort currently required for the measurement of the latter groups, a tailored methodology is presented that is capable of providing valuable process insights as well as predicting the glycosylation profile based on only four experiments measured over 12 cell culture days. An important result of the work is the possibility to accurately predict many of the glycan variables based on the information of three experiments. An additional finding is that such predictive models can be generated from the more accessible process and extracellular information only, that is, without including the more experimentally cumbersome intracellular data. With regards to the incorporation of omics data in the standard process analytics framework in the future, this works provides a comprehensive data analysis pathway which can efficiently support numerous bioprocessing tasks.

Hybrid-EKF: Hybrid model coupled with extended Kalman filter for real-time monitoring and control of mammalian cell culture.

In a decade when Industry 4.0 and quality by design are major technology drivers of biopharma, automated and adaptive process monitoring and control are inevitable requirements and model-based solutions are key enablers in fulfilling these goals. Despite strong advancement in process digitalization, in most cases, the generated datasets are not sufficient for relying on purely data-driven methods, whereas the underlying complex bioprocesses are still not completely understood. In this regard, hybrid models are emerging as a timely pragmatic solution to synergistically combine available process data and mechanistic understanding. In this study, we show a novel application of the hybrid-EKF framework, that is, hybrid models coupled with an extended Kalman filter for real-time monitoring, control, and automated decision-making in mammalian cell culture processing. We show that, in the considered application, the predictive monitoring accuracy of such a framework improves by at least 35% when developed with hybrid models with respect to industrial benchmark tools based on PLS models. In addition, we also highlight the advantages of this approach in industrial applications related to conditional process feeding and process monitoring. With regard to the latter, for an industrial use case, we demonstrate that the application of hybrid-EKF as a soft sensor for titer shows a 50% improvement in prediction accuracy compared with state-of-the-art soft sensor tools.

Process-wide control and automation of an integrated continuous manufacturing platform for antibodies.

Integrated continuous manufacturing is entering the biopharmaceutical industry. The main drivers range from improved economics, manufacturing flexibility, and more consistent product quality. However, studies on fully integrated production platforms have been limited due to the higher degree of system complexity, limited process information, disturbance, and drift sensitivity, as well as difficulties in digital process integration. In this study, we present an automated end-to-end integrated process consisting of a perfusion bioreactor, CaptureSMB, virus inactivation (VI), and two polishing steps to produce an antibody from an instable cell line. A supervisory control and data acquisition (SCADA) system was developed, which digitally integrates unit operations and analyzers, collects and centrally stores all process data, and allows process-wide monitoring and control. The integrated system consisting of bioreactor and capture step was operated initially for 4 days, after which the full end-to-end integrated run with no interruption lasted for 10 days. In response to decreasing cell-specific productivity, the supervisory control adjusted the loading duration of the capture step to obtain high capacity utilization without yield loss and constant antibody quantity for subsequent operations. Moreover, the SCADA system coordinated VI neutralization and discharge to enable constant loading conditions on the polishing unit. Lastly, the polishing was sufficiently robust to cope with significantly increased aggregate levels induced on purpose during virus inactivation. It is demonstrated that despite significant process disturbances and drifts, a robust process design and the supervisory control enabled constant (optimum) process performance and consistent product quality.

Bioprocessing in the Digital Age: The Role of Process Models.

In this age of technology, the vision of manufacturing industries built of smart factories is not a farfetched future. As a prerequisite for Industry 4.0, industrial sectors are moving towards digitalization and automation. Despite its tremendous growth reaching a sales value of worth $188 billion in 2017, the biopharmaceutical sector distinctly lags in this transition. Currently, the challenges are innovative market disruptions such as personalized medicine as well as increasing commercial pressure for faster and cheaper product manufacturing. Improvements in digitalization and data analytics have been identified as key strategic activities for the next years to face these challenges. Alongside, there is an emphasis by the regulatory authorities on the use of advanced technologies, proclaimed through initiatives such as Quality by Design (QbD) and Process Analytical Technology (PAT). In the manufacturing sector, the biopharmaceutical domain features some of the most complex and least understood processes. Thereby, process models that can transform process data into more valuable information, guide decision-making, and support the creation of digital and automated technologies are key enablers. This review summarizes the current state of model-based methods in different bioprocess related applications and presents the corresponding future vision for the biopharmaceutical industry to achieve the goals of Industry 4.0 while meeting the regulatory requirements.

A new generation of predictive models: The added value of hybrid models for manufacturing processes of therapeutic proteins.

Due to the lack of complete understanding of metabolic networks and reaction pathways, establishing a universal mechanistic model for mammalian cell culture processes remains a challenge. Contrarily, data-driven approaches for modeling these processes lack extrapolation capabilities. Hybrid modeling is a technique that exploits the synergy between the two modeling methods. Although mammalian cell cultures are among the most relevant processes in biotechnology and indeed looks ideal for hybrid modeling, their application has only been proposed but never developed in the literature. This study provides a quantitative assessment of the improvement brought by hybrid models with respect to the state-of-the-art statistical predictive models in the context of therapeutic protein production. This is illustrated using a dataset obtained from a 3.5 L fed-batch experiment. With the goal to robustly define the process design space, hybrid models reveal a superior capability to predict the time evolution of different process variables using only the initial and process conditions in comparison to the statistical models. Hybrid models not only feature more accurate prediction results but also demonstrate better robustness and extrapolation capabilities. For the future application, this study highlights the added value of hybrid modeling for model-based process optimization and design of experiments.

A new flow cell and chemometric protocol for implementing in-line Raman spectroscopy in chromatography.

On-line monitoring tools for downstream chromatographic processing (DSP) of biotherapeutics can enable fast actions to correct for disturbances in the upstream, gain process understanding, and eventually lead to process optimization. While UV/Vis spectroscopy is mostly assessing the protein's amino acid composition and the application of Fourier transform infrared spectroscopy is limited due to strong water interactions, Raman spectroscopy is able to assess the secondary and tertiary protein structure without significant water interactions. The aim of this work is to implement the Raman technology in DSP, by designing an in-line flow cell with a reduced dead volume of 80 µL and a reflector to increase the signal intensity as well as developing a chemometric modeling path. In this context, measurement settings were adjusted and spectra were taken from different chromatographic breakthrough curves of IgG1 in harvest. The resulting models show a small average RMSEP of 0.12 mg/mL, on a broad calibration range from 0 to 2.82 mg/mL IgG1. This work highlights the benefits of model assisted Raman spectroscopy in chromatography with complex backgrounds, lays the fundamentals for in-line monitoring of IgG1, and enables advanced control strategies. Moreover, the approach might be extended to further critical quality attributes like aggregates or could be transferred to other process steps.

Decision Tree-PLS (DT-PLS) algorithm for the development of process: Specific local prediction models.

This work presents a novel multivariate statistical algorithm, Decision Tree-PLS (DT-PLS), to improve the prediction and understanding of dynamic processes based on local partial least square regression (PLSR) models for characteristic process groups defined based on Decision Tree (DT) analysis. The DT-PLS algorithm is successfully applied to two different cell culture data sets, one obtained from bioreactors of 3.5 L lab scale and the other obtained from the 15 ml ambr microbioreactor system. Substantial improvement in the predictive capabilities of the model can be achieved based on the localization compared to the classical PLSR approach, which is implemented in the commercially available packages. Additionally, the differences in the model parameters of the local models suggest that the governing process variables vary for the different process regimes indicating the different states of the cell under different process conditions.

Sequential Multivariate Cell Culture Modeling at Multiple Scales Supports Systematic Shaping of a Monoclonal Antibody Toward a Quality Target.

The development of cell culture processes is highly complex and requires a large number of experiments on various scales to define the design space of the final process and fulfil the regulatory requirements. This work follows an almost complete process development cycle for a biosimilar monoclonal antibody, from high throughput screening and optimization to scale-up and process validation. The key goal of this analysis is to apply tailored multivariate tools to support decision-making at every stage of process development. A toolset mainly based on Principal Component Analysis, Decision Trees, and Partial Least Square Regression combined with a Genetic Algorithm is presented. It enables to visualize the sequential improvement of the high-dimensional quality profile towards the target, provides a solid basis for the selection of effective process variables and allows to dynamically predict the complete set of product quality attributes. Additionally, this work shows the deep level of process knowledge which can be deduced from small scale experiments through such multivariate tools. The presented methodologies are generally applicable across various processes and substantially reduce the complexity, experimental effort as well as the costs and time of process development.

Enhanced process understanding and multivariate prediction of the relationship between cell culture process and monoclonal antibody quality.

This work investigates the insights and understanding which can be deduced from predictive process models for the product quality of a monoclonal antibody based on designed high-throughput cell culture experiments performed at milliliter (ambr-15® ) scale. The investigated process conditions include various media supplements as well as pH and temperature shifts applied during the process. First, principal component analysis (PCA) is used to show the strong correlation characteristics among the product quality attributes including aggregates, fragments, charge variants, and glycans. Then, partial least square regression (PLS1 and PLS2) is applied to predict the product quality variables based on process information (one by one or simultaneously). The comparison of those two modeling techniques shows that a single (PLS2) model is capable of revealing the interrelationship of the process characteristics to the large set product quality variables. In order to show the dynamic evolution of the process predictability separate models are defined at different time points showing that several product quality attributes are mainly driven by the media composition and, hence, can be decently predicted from early on in the process, while others are strongly affected by process parameter changes during the process. Finally, by coupling the PLS2 models with a genetic algorithm first the model performance can be further improved and, most importantly, the interpretation of the large-dimensioned process-product-interrelationship can be significantly simplified. The generally applicable toolset presented in this case study provides a solid basis for decision making and process optimization throughout process development. © 2017 American Institute of Chemical Engineers Biotechnol. Prog., 33:1368-1380, 2017.

Parallel experimental design and multivariate analysis provides efficient screening of cell culture media supplements to improve biosimilar product quality.

Rational and high-throughput optimization of mammalian cell culture media has a great potential to modulate recombinant protein product quality. We present a process design method based on parallel design-of-experiment (DoE) of CHO fed-batch cultures in 96-deepwell plates to modulate monoclonal antibody (mAb) glycosylation using medium supplements. To reduce the risk of losing valuable information in an intricate joint screening, 17 compounds were separated into five different groups, considering their mode of biological action. The concentration ranges of the medium supplements were defined according to information encountered in the literature and in-house experience. The screening experiments produced wide glycosylation pattern ranges. Multivariate analysis including principal component analysis and decision trees was used to select the best performing glycosylation modulators. Subsequent D-optimal quadratic design with four factors (three promising compounds and temperature shift) in shake tubes confirmed the outcome of the selection process and provided a solid basis for sequential process development at a larger scale. The glycosylation profile with respect to the specifications for biosimilarity was greatly improved in shake tube experiments: 75% of the conditions were equally close or closer to the specifications for biosimilarity than the best 25% in 96-deepwell plates. Biotechnol. Bioeng. 2017;114: 1448-1458. © 2017 Wiley Periodicals, Inc.

Robust factor selection in early cell culture process development for the production of a biosimilar monoclonal antibody.

This work presents a multivariate methodology combining principal component analysis, the Mahalanobis distance and decision trees for the selection of process factors and their levels in early process development of generic molecules. It is applied to a high throughput study testing more than 200 conditions for the production of a biosimilar monoclonal antibody at microliter scale. The methodology provides the most important selection criteria for the process design in order to improve product quality towards the quality attributes of the originator molecule. Robustness of the selections is ensured by cross-validation of each analysis step. The concluded selections are then successfully validated with an external data set. Finally, the results are compared to those obtained with a widely used software revealing similarities and clear advantages of the presented methodology. © 2016 American Institute of Chemical Engineers Biotechnol. Prog., 33:181-191, 2017.

Fingerprint detection and process prediction by multivariate analysis of fed-batch monoclonal antibody cell culture data.

This work presents a sequential data analysis path, which was successfully applied to identify important patterns (fingerprints) in mammalian cell culture process data regarding process variables, time evolution and process response. The data set incorporates 116 fed-batch cultivation experiments for the production of a Fc-Fusion protein. Having precharacterized the evolutions of the investigated variables and manipulated parameters with univariate analysis, principal component analysis (PCA) and partial least squares regression (PLSR) are used for further investigation. The first major objective is to capture and understand the interaction structure and dynamic behavior of the process variables and the titer (process response) using different models. The second major objective is to evaluate those models regarding their capability to characterize and predict the titer production. Moreover, the effects of data unfolding, imputation of missing data, phase separation, and variable transformation on the performance of the models are evaluated.

Model-based description of peptide retention on doped reversed-phase media.

Reversed-phase (RP) chromatography is one of the main tools for the preparative purification of therapeutic peptides. In previous works [1,2], a new type of RP chromatography, doped reversed-phase chromatography (DRP) was presented. By adding small amounts (up to 15% of the surface ligands) of repulsive ion exchange ligands to a traditional RP material, significant improvements in peptide purification performance were observed, at the same or in similar operating conditions. These improvements included increases in selectivity in diluted conditions (up to twice as high), increases in yield in preparative conditions (up to 20% higher) and in productivity in preparative conditions (up to twice as high), when compared to RP materials [2]. A proper physical model is developed in this work to quantitatively explain and rationalize this behavior. The developed model is then used to correctly fit the retention data of several peptides in different buffering conditions. The increase in selectivity is related to a controlled decrease in free surface area available for adsorption due to the ionic ligands creating a repulsive sphere the analytes cannot enter. This decrease in adsorption surface is calculated using Debye-Hückel theory, and in combination with linear solvent strength theory, allows for the quantitative description of peptide retention on DRP media.

Doping reversed-phase media for improved peptide purification.

The purification of therapeutic peptides is most often performed using one or more reversed phase chromatography steps. This ensures high purities while keeping the costs of purification under control. In this paper, a doped reversed phase chromatographic material is tested and compared to traditional reversed phase materials. The doping consists of adding limited amounts of ion exchange ligands to the surface of the material to achieve orthogonal separation and increase the non-hydrophobic interactions with the surface. These ionic groups can either be attractive (opposite charge), or repulsive (same charge) to the peptide. The benefit of this new doped reversed phase material is shown through increases in selectivity in diluted conditions and yield and productivity in overloaded (i.e. industrial) conditions. It is the conjectured that all performance characteristics should increase using repulsive doping groups, whereas these characteristics should decrease when using attractive doping groups. This conjecture is shown to be true through several examples, including purifications of industrially relevant peptide crudes, in industrially relevant conditions. Moreover, the effect of ionic strength and organic modifier concentration was explored and shown to be in line with the expected behavior.