The safety requirements for vaccines are very high because they are intended for healthy people. Therefore, vaccine development is time-consuming and very expensive. Reducing time to market is key for pharmaceutical companies, saving lives and money. Therefore, systematic, versatile, and effective process development strategies are needed to shorten development time and enhance process understanding. High-throughput technologies have greatly increased the amount of useful data related to processes, and new high-throughput process development (HTPD) methods are constantly being developed, combining statistical and mechanistic modeling. The introduction of model-based HTPD enables faster and broader screening of conditions and further increases knowledge. Model-based HTPD is particularly important for chromatography , which is a key separation technology to achieve high purity.
This review will introduce the downstream process development strategies and tools used in the biopharmaceutical industry, focusing on the vaccine purification process. High-throughput process development and other combinatorial approaches will then be discussed and compared based on experimental work and understanding. In an ever-growing ocean of information, new modeling tools and artificial intelligence (AI) are becoming increasingly important for discovering patterns behind the data and therefore gaining a deeper process understanding.
[ Taking vaccines as an example, review the latest strategic progress in accelerating purification process development ]
Downstream process development methods
The overall goal of process development is to design the optimal purification process and strive to achieve the purity target by minimizing cost and time while complying with all regulatory requirements. Currently, vaccine development mainly employs DoE-based approaches, but it could also benefit from more advanced, model-based process development approaches that are already used in other biopharmaceutical areas, such as the purification of mAbs. Figure 3 shows two types of process development approaches, DoE-based approaches and modeling-based approaches. In the following, the process development methodology is briefly described.
Figure 3. Overview of two different process development approaches. Left: Design of Experiments (DoE) approach, which conducts experiments based on statistical tools and evaluates results through statistical analysis. This approach is commonly used in the biopharmaceutical industry. Right: Model-based process development approach, where model input parameters, such as isotherm parameters and column parameters, are determined through targeted experiments. Before performing optimization, the model must be validated.
Experiment-driven downstream process development
One factor at a time (OFAT) and design of experiments (DOE)
One factor at a time (OFAT) is a more traditional approach in which one factor is changed during a series of experiments while the other factors are held constant. In this method, the dependence between factors is ignored, so finding the optimal value is quite difficult and inefficient. For this reason, more than a decade ago, the biopharmaceutical industry turned to statistical-based DoE methods to design and analyze experiments to obtain more valuable information through fewer experiments. The classic DoE method is factorial design. Experiments are conducted on all possible combinations of factors with the goal of determining the impact of each factor and the impact of interactions between factors on the response. An improvement over classic DoE screening is a deterministic screening design that estimates curvature effects and is able to distinguish factors that have a significant impact on the response from factors that have a negligible impact. Other methods that provide a three-level multifactor design are the Box-Behnken or central composite designs. Hibbert D.B. provides an in-depth introduction to the most commonly used DoE methods, focusing on over . There are various DoE software such as Design-Expert, Modde and JMP, but other statistical software such as R, SPSS and various Python packages can also be used for DoE purposes.
Data acquisition for modeling purposes
Another experimental strategy is to determine parameters that serve as input to a mechanical or physical model. The use of mechanical models was established decades ago and is now widely adopted by the chemical industry, with some processes even designed entirely on computers.Only recently did the biopharmaceutical and vaccine industries initiate this strategy in their process development, where the main challenge was often complex injection mixtures containing the product of interest (e.g. antigen ) as well as thousands of proteins and impurities. This is probably why mechanistic modeling and parameter acquisition have not been widely adopted, as it is almost impossible to experimentally determine and model thousands of proteins and impurities. However, HTE makes determining model parameters well worth it, even for more complex mixtures. Notably, validated models increase understanding of the process and enable it to be optimized in silico, saving time, material and cost. For chromatography purposes, as this is the main purification technique in protein subunit vaccines, adsorption isotherm parameters describing the binding behavior of the components to the solid phase are crucial. Experimental determination of adsorption equilibrium is necessary to establish isotherm parameters and can be obtained by batch adsorption experiments, frontal analysis, isocratic or linear gradient elution, or by utilizing inverse techniques, by adjusting certain parameters to minimize the difference between experimental and simulated elution profiles. In addition to isotherm determination, chromatography columns and packing materials must be characterized to obtain a validated model, but these are more readily available.
High-throughput screening
About two decades ago, the execution of experiments was accelerated by the introduction of liquid handling stations (LHS), also known as high-throughput experiments (THE) or high-throughput screening (HTS). Due to automation, miniaturization and parallelization, the creation of large data sets becomes feasible while allowing for shorter time periods using smaller sample sizes and resources. Another benefit of automation is that it reduces variability and provides superior accuracy. Today, LHS has become a widely used technology in academia and industry, significantly reducing process development time. Since LHS allows more conditions to be screened, it is more feasible to find the optimal conditions for the purification process. In addition to the benefits of the system, some disadvantages have also been reported in the literature. For example, LHS has limitations in accurately modeling flow distribution in process columns. HTS requires a high level of understanding of effective experimental design to optimally exploit the system, so it is more of a tool than a stand-alone method.
Expertise-driven downstream process development
General experience
Rules of thumb, available knowledge and experience of existing processes are the basis for expertise or heuristics in designing new production processes. Expertise insights are easily applied and can accelerate process design by eliminating combinations of unit operations with less than ideal outcomes. Asenjo et al. developed a specialized system focused on downstream protein processing; the software uses a database of common process design (heuristics) expertise to support and accelerate decisions in selecting a series of unit operations. Some manuals also provide broad overviews of general design heuristics. Most vaccine purification processes are also based on heuristic methods, such as the purification of hepatitis A virus from mammalian cell culture, where the first step involves low-cost anion exchange chromatography to capture the product and remove large amounts of impurities, and the final steps of the downstream process are purification and desalting steps using size exclusion chromatography. A general example that is almost entirely knowledge-based is platform crafting, which is explained in more detail in the next paragraph. The
platform process
platform process can be used as a "template" to design an entire purification sequence for a specific type of molecule using a pre-established series of unit operations. The platform description provides detailed information on the operating conditions for each unit operation, corresponding to the entire purification process. One of the key advantages is the reduction in process development time, associated regulatory and resource requirements for similar molecules, thereby reducing time to market and validation efforts. In addition, platform files can be shared and matched not only between different departments, but also between different production sites, becoming a site-independent process. Platform process methods are best suited for biopharmaceuticals with similar characteristics and therefore similar purification steps. For example, mAbs are relatively well defined, and platform processes can be used to establish similar purification procedures for new mAb isoforms.Detailed information on process-related contaminants such as HCP and other impurities of corresponding cell cultures, i.e. CHO and hybridoma cells, is known. The sequence of purification steps includes Protein A chromatography, low pH virus inactivation, IEX chromatography purification step, virus removal filtration, and ultrafiltration/diafiltration. The purification process for new mAb isoforms can be determined with only minor changes to the purification process conditions. Other potentially applicable product candidates for the platform approach could be pDNA vaccines and influenza vaccines, both of which have similar properties and purification steps. However, the properties of mAbs are relatively similar, while the appearance of protein subunit vaccines varies greatly, making the purification process more difficult to standardize.
Model-based downstream process development
In process engineering, models play an important role, they are designed to represent a real system in an abstract mathematical form. Bézivin and Gerbé define a model as "a simplification of a system built with the intended goals in mind. The model should be able to answer questions in place of the actual system". The intended goals related to process engineering can be control, simulation, design, monitoring or optimization. Depending on the goal, different models may apply. Models help understand complex problems and can provide potential solutions if they adequately represent the target characteristics of the modeled system. Running a model with a given set of parameters is called a simulation, and is a cheap and safe way to run virtual experiments. Therefore, the number of experiments in the laboratory can be reduced and/or designed more efficiently, thereby reducing time and material consumption. Although using models sounds attractive and promising, developing sound models that achieve the intended purpose requires time, effort, and knowledge. Additionally, there is a shortage of people in the field who can develop and maintain scientific and engineering software. In the near future, it is expected that more process engineers or scientists will become familiar with modeling, as most technology-related studies now offer courses in programming and data processing. In order to build a model, two main resources are essential, namely the knowledge of the process, translated into natural laws, and the collection of data obtained from real systems. In process engineering, a distinction can be made between first principles, mechanical or knowledge-driven models and data-driven or empirical models, known respectively as transparent white-box models and less transparent black-box models. The combination of the two is called a hybrid semi-parametric model. Table 1 outlines the main advantages and disadvantages.
Data-driven models
Data-driven or empirical models attempt to describe input-output relationships based on observed experiments within a predefined design space, such as Artificial Neural Networks (ANN), statistical and regression models. The biopharmaceutical industry often uses statistical models, either by performing a set of predefined experiments using DoE and appropriate statistical data analysis methods such as response surface methods (RSM), or by using existing data sets for multivariate data analysis. RSM is a well-known empirical model that describes the response relationships between different test factors within the DoE and generates a model describing the mathematical relationships. This statistical ( black box ) model only observes the correlation of factors with responses without obtaining a basic mechanistic (physicochemical) understanding of the estimated parameters. Using DoE and regression analysis through first- and second-order polynomials , the optimal input combination can be estimated. However, fitting the data to a second-order polynomial is a major drawback of RSM, since typically not all curvature within a system can be described by a second-order polynomial. DoE combined with empirical models has been widely used in downstream purification process design in the biopharmaceutical industry and academia. Some researchers used DoE and linear regression model to study the effect of high-salt solution on RNA precipitation and pDNA recovery. Recently, Chiang et al. used DoE to evaluate the impact of chromatography parameters on virus clearance when switching from single-column operation to multi-column operation. A major limitation of data-driven models is that they are only valid within a defined region of the measured variable and can only predict variables within that region, making extrapolation often very inaccurate. Furthermore, little process knowledge can be extracted since the parameters are often only correlations.On the other hand, data-driven modeling does not require advance knowledge of the process and is less time-consuming than mechanical modeling.
Mechanical Modeling
Mechanical, first principles, or knowledge-driven models attempt to describe the internal mechanisms and phenomena occurring in a process or system based on knowledge about the process. These models consist of material and/or energy balances as well as transport and thermodynamics phenomena and have a fixed structure, which means that the parameters may have physical interpretations. Model parameters are estimated from experimental data or physical correlations. The physical processes that occur during purification can be translated into mathematical simulation models. A validated mechanical model allows various conditions to be explored in silico and therefore optimal operating conditions can be obtained efficiently. The phenomena occurring within the chromatography column are well described in the literature, and Ruthven has covered the kinetics and adsorption processes in depth. Kinetic or rate models are most common in practice and include dispersion factors, such as mass transfer and dispersion effects, and equilibrium factors, such as adsorption isotherms, ion dissociation, and intermolecular associations. The three most prominent kinetic models are the lumped kinetic model, the lumped pore model, and the general rate model, arranged in order of complexity. The main difference between these models is the extent to which pore diffusion effects are covered. However, it is applicable to all mechanical models where the isotherm parameters are crucial and, as mentioned earlier, there are many combined models such as linear, Langmuir, spatial mass action, and mixed modes. The use of chromatography models ranges from process synthesis, optimization and control to scale-up, filler selection and robustness studies. A further step is to simulate combinations of integrated chromatography and other conditioning steps to find the optimal overall purification process. There are currently a variety of commercial software for chromatography mechanical models available, such as: GoSilico (a subsidiary of Cytivia, officially known as ChromX), Aspen Chromatography, DelftChrom, CADET and ChromaTech.
In silico alternatives for adsorption experiments have been studied for several years. Molecular dynamics simulations attempt to describe filler-protein interactions at a detailed atomic level. Quantitative structure-activity relationships (QSAR) combine molecular properties with empirical models to discover correlations between retention behavior and protein surface properties. This molecular model can be used to predict protein retention behavior on fillers to reduce process development time. However, detailed information about each component, such as the amino acid sequence or the crystal structure , is usually required, and extensive experimentation is also required.
Compared with data-driven models, mechanistic models can explore a wider range of conditions even beyond the observed measurements, with higher extrapolation capabilities. This facilitates process understanding, in line with QbD initiatives, although mechanical modeling also requires physical understanding. The main disadvantage of knowledge-based models is their complexity and thus require more development time compared to data-driven models.
Hybrid (semi-parametric) modeling
Hybrid (semi-parametric) modeling combines parametric (i.e., first principles, mechanical, and knowledge-based models) with non-parametric (i.e., data-driven models) to eliminate the shortcomings of individual methods and get the best of both. Von Stosch et al. provide an extensive review of hybrid semiparametric modeling frameworks and various applications in biochemistry engineering regarding process monitoring, control, optimization, model reduction, and scale-up. Parametric and nonparametric models can be configured in series or parallel, depending on the scope of the model. Parallel mode is often recommended when mechanical (white-box) model performance is limited or inaccurate enough, and adding a non-parametric (black-box) model may improve the evaluation, as shown in Figure 4c. Serial methods are often used to reduce the complexity of a mechanical model by determining parameters using a non-parametric model, as in Figure 4a, or when the results of a mechanical model are used as input to a non-parametric model, as in Figure 4b.
Figure 4. Hybrid modeling configuration, white box represents mechanical/first principles model, black box represents data-driven model. Serial method (A, B) and parallel method (C).
The usefulness of hybrid modeling lies in its ability to cost-effectively solve complex problems and develop models.In addition to gaining broader process understanding, other benefits include optimized model accuracy, transparency and extrapolation properties. The challenge, however, is to understand in what way different types of models can be combined to develop hybrid models. Therefore, a thorough understanding of data-driven and mechanistic models is required, as well as the knowledge to obtain the correct data. Hybrid modeling is receiving increasing attention from industry and academia and appears to be a promising approach to overcome the shortcomings of data-driven and mechanistic models.
This article is excerpted and translated from the original text below. Due to limited level, please refer to the original text for details. This article is intended to share knowledge and information. If you have any questions, please contact us via private message.
Original article: D.Keulen, G.Geldhof, O.L.Bussy, et al., Recent advances to accelerate purification process development: a review with a focus on vaccines. Journal of Chromatography A, 2022, https://doi.org/10.1016/j.chroma.2022.463195.
