Downstream process development strategies for effective bioprocesses: Trends, progress, and combinatorial approaches

Abstract

The biopharmaceutical industry is at a turning point moving toward a more customized and patient‐oriented medicine (precision medicine). Straightforward routines such as the antibody platform process are extended to production processes for a new portfolio of molecules. As a consequence, individual and tailored productions require generic approaches for a fast and dedicated purification process development. In this article, different effective strategies in biopharmaceutical purification process development are reviewed that can analogously be used for the new generation of antibodies. Conventional approaches based on heuristics and high‐throughput process development are discussed and compared to modern technologies such as multivariate calibration and mechanistic modeling tools. Such approaches constitute a good foundation for fast and effective process development for new products and processes, but their full potential becomes obvious in a correlated combination. Thus, different combinatorial approaches are presented, which might become future directions in the biopharmaceutical industry.

Abbreviations

AEX

anion‐exchange

ambr

advanced micro‐scale bioreactor

CEX

cation‐exchange

CHO

Chinese hamster ovary

DoE

design of experiments

DSP

downstream processing

EDM

equilibrium dispersive model

GRM

general rate model

HIC

hydrophobic interaction chromatography

HT

high‐throughput

HTPD

HT process development

IB

inclusion body

IEX

ion‐exchange

MCM

multivariate calibration modeling

MTP

microtiter plate

QSAR

quantitative structure‐activity relationship

QSPR

quantitative structure‐property relationship

RSM

response surface modeling

USP

upstream processing

1. Introduction

The biopharmaceutical industry is rapidly changing and the production portfolio is steadily expanded for new biologically active molecules. The production of pharmaceutically relevant agents covers small molecules as well as complex biomolecules such as mAbs 1, protein vaccines 2, enzymes, hormones, and growth factors 3. As new promising biomolecules are steadily identified, there is a need for fast and reliable process development and optimization tools. This includes extending the large‐quantity blockbuster production of pharmaceuticals (one‐drug‐fits‐all approach) with custom designs 4. Especially, the product purification process can be challenging for meeting the requirements stated by the authorities such as the Food and Drug Administration. For meeting the given regulatory standards, new technologies in protein purification in terms of strategies, materials, and additional statistical and in silico methods need to be considered.

The optimization and combination of different unit operations and modes of operation create a need for advanced process characterization tools for generating deeper process understanding and exploring the experimental design space. Nfor et al. 5 discussed downstream process optimization strategies within three distinct categories, which are expert‐, experimental‐, and model‐based approaches. This categorization is a good basis for process optimization and was also described by Guiochon and Beaver 6. All three approaches are highly useful for effective process development for new products, and their full potential can be realized in a combination of these procedures in so‐called combinatorial approaches.

Figure 1 gives an overview of different methodologies for process development and optimization and their combination. The heuristic approach based on expert knowledge and rules of thumb can be used for the generation of initial datasets for the high‐throughput (HT) experimental approach for screening in the feasible region of the design space. Inversely, the experimental approach is used for characterizing processes and generating new platforms and databases. When defined HT experiments are used for lysate characterization of unknown compositions and a systematic prediction of key parameters for new tailored processes, this can be referred to as a deterministic approach 7.

Figure 1.

Schematic illustration of different tools for downstream process development. Heuristic, experimental, and model‐based approaches form the foundation of process design alternatives. Combinatorial approaches such as deterministic and hybrid procedures represent new synergetic strategies.

When moving toward in silico methods for process characterization, the experimental approach is used for model calibration and validation. The model‐based approach can be based on both, empirical models such as response surface modeling (RSM) or mechanistic models. Both types of models are used for computational optimization and reduction of experimental effort. A combination of HT experiments for generating different feedstocks and experimental determination of model parameters, and in silico methods is referred to as a hybrid approach 8, 9.

2. Heuristic approaches

Heuristic approaches exploit expert knowledge and rules of thumb to forecast the behavior of a new product or production path based on analogous previous experiences. Expert knowledge can be subdivided into three different categories: universal process design heuristics, heuristics for specified processes, and platform processes.

2.1. Universal process design heuristics

Universal process design heuristics are general rules of thumb for process design and were described in diverse databases called expert systems 10, 11. A list of different rules for chemical engineering processes was described by Nishida et al. 12. Most of these suggestions can be applied accordingly for bioprocesses including basic rules such as:

1. Perform the most complicated separation at the end (e.g. removal of product‐related impurities such as product aggregates and fragments)

2. Remove the impurities of highest quantity first (e.g. cell debris and water)

3. Remove destabilizing and unstable material as early as possible within the process (e.g. proteases)

Wheelwright 13 and Asenjo and Patrick 14 appended further points for biopharmaceutical separation processes. One very important suggestion is the use of orthogonal purification strategies during process design for exploiting different physicochemical properties of molecules during each stage. An anion‐exchange (AEX) step followed by a cation‐exchange (CEX) separation is thus superior to a consecutive use of two separation steps of the same type. Also, each unit operation must exploit the maximum physicochemical differences of products and impurities. It is obvious that such expert advices require deeper knowledge of the feed composition and molecule properties. Further universal design heuristics include using the most product‐selective process step at the early stage of a process to minimize the total number of purification stages while performing the most expensive step at the final stage 14. Guidelines found in the Recombinant Protein Purification Handbook by GE Healthcare Life Sciences 15 and the Handbook of Process Chromatography by Sofer and Hagel 16 are even more specific:

• (iv)The combination of unit operations needs to be in an order to minimize the number of conditioning steps
• (v)Therapeutic proteins generally require viral removal steps
• (vi)Use SEC in final stages only
• (vii)If there is no knowledge about the product and the complex feed mixture, a combination of ion‐exchange (IEX) chromatography, hydrophobic interaction chromatography (HIC), and SEC is a good start.

As universal design heuristics are rather straightforward, they are found in almost all existing workflows in biopharmaceutical productions. A good example is the hepatitis A virus vaccine production purification process from mammalian cells 17. Product capturing is performed by rather selective and cheap AEX chromatography and precipitation steps followed by SEC polishing.

However, the platform process for mAbs is the perfect example of how expert knowledge can be controversial. The protein A affinity chromatography step is both, the most expensive step and should thus be performed last, and the step of highest resolution which is why it should be conducted at the beginning.

2.2. Heuristics for specified processes

Besides the universal process design heuristics discussed in 2.1, expert knowledge is also present for single process steps. These rules of thumb can help optimizing chosen unit operations within a fixed process.

For HIC, protein precipitation and protein unfolding, the Hofmeister (or lyotropic) series of salts has been used for over a century 18. The listed salts are sorted in the order of their chaotropic (lowering the structure of water) and kosmotropic (strengthening the structure of water) nature 19. Kosmotropic salts like ammonium sulfate are generally regarded as enhancers for hydrophobic forces and are commonly used for strengthening HIC‐binding effects 20, 21, 22, 23 and protein precipitation 24. The cavity theory is an extension of the Hofmeister theory and was described by Melander and Horváth 25 as well as Senczuk et al. 26 who say that salts creating higher surface tension (as most kosmotropic salts) promote HIC binding.

By contrast, chaotropic agents such as guanidium chloride are widely applied as solvating agents as in inclusion body (IB) unfolding applications 27, 28. Whereas these rules are consistent in most cases, they often fail when synergetic effects come into play. Mixed salts of chaotropic and kosmotropic nature were reported to enhance hydrophobicity to a larger extent than the same amount of kosmotropic salt alone 29, 30.

Further process‐related heuristics can be found for IEX chromatography unit operations. A common rule for IEX process development is the binding of the complex mixture 0.5 pH units above (AEX chromatography) or below (CEX chromatography) the pI of the product 31. The idea of this rule of thumb is to ensure a sufficient charge for the binding of the product while keeping most of the impurities in the flow‐through fraction 32, 33. As the pI of a protein accounts for overall net charge only, such an approach fails when extreme surface charge distributions on the protein surface come into play 34. Such an approach also fails when different glycosylation patterns 35 and additional hydrophobic interactions with the adsorbent backbone come into play 36.

For HIC processes that are performed at pH values in the neutral region 37, 38, a similar rule of thumb was proposed 39. Instead of binding in the standard pH range of 7–8, pH values close to the pI of the protein and, generally, regions of reduced solubility are good starting points for HIC process optimization. Such behavior was encountered for proteins of acidic, neutral, and alkaline pI.

2.3. Platform processes

Platform processes are a strict series of unit operations that can be used as an entire purification sequence for a certain type of molecule. Such template processes, which proved to be successful for similar applications, resemble an exception of expert‐based approaches. Platform processes can directly be transferred and are in most cases not questioned for alternatives. Advantages of platform processes are the reduction in time until clinical trials (template processes are acknowledged by authorities) as well as the use of established knowledge and process understanding created in former applications 40.

The platform process most applied in biopharmaceutical downstream processing (DSP) is the systematic purification of mAbs and Fc fusion proteins. Shukla et al. 41 described the antibody platform process in several consecutive steps, which are consistent in process structure, but single unit operations are exchangeable. After cell harvest and solid–liquid separation, the supernatant is applied to a high‐affinity protein A chromatography column for selective product isolation. The further process contains two different virus inactivation and removal steps as well as two orthogonal chromatography polishing steps for removal of process‐ (DNA, host cell proteins, etc.) and product‐related impurities (high‐ and low‐molecular weight species, charge and glycosylation variants). In a consecutive diafiltration step, the product is stabilized and prepared for the final formulation. Different pharmaceutical companies adapted this template process and until now, most licensed processes for mAbs are based on a similar scheme. Genentech uses protein A chromatography as a capturing step followed by a sequence of CEX and AEX chromatography for polishing 42. Shukla et al. 41 summarized unit operations used for mAb polishing steps within the platform process classified for different impurities including AEX, CEX, HIC, and hydroxylapatite chromatography.

3. Experimental approaches

Expert knowledge and platform processes are only applicable for a limited range of products and processes. Designing tailored processes for new biological and expression systems can hardly be realized based on heuristics only. When there is no knowledge about a certain molecule or process, the experimental approach in terms of high‐throughput process development (HTPD) is a straightforward solution for a broad exploration of the design space.

HTPD as a methodology for a large number of parallelized, miniaturized, and automated experiments is one of the hot topics in pharmaceutical research and evolves to a standard tool in industry 43, 44. The large number of review articles and journal special issues underline the need for such methodologies for biopharmaceutical process development 43, 45, 46. In this section, miniaturized systems at different stages of biopharmaceutical process development are reviewed.

3.1. HT cultivation systems

Cell cultivations and product formation under a vast number of parameters and process conditions strongly influence the product purification process 47. The screening for optimal expression systems, media, additives, process conditions, and induction setups for up‐ and downstream applications explains the need for cultivation systems in miniaturized format. The first overview of different commercial micro‐cultivation systems and their specifications are shown in Table 1. While such HT systems are operated with small‐volume reactors, many of these systems lack in process control compared to fully equipped bioreactors 48. Cultivations in standard 96‐well microtiter plates (MTPs) represent the optimal system for miniaturization while reducing the process information content to a minimum 49. Ninety‐six well approaches in standard MTPs were used for the rapamycin production in Streptomyces hygroscopius 50, Glucansucrase from Leuconostoc mesenteroides mutants 51, and recombinant human protein 1 from Chinese hamster ovary (CHO) cells 52.

Table 1.

Overview of different commercial miniaturized cultivation systems and their application

System	Distributor	Reactors	Working volume (mL)	PAT tools	Reactor type	Applications
ambr	Tap Biosystems	up to 48	10–15	pH DOT	STR	mAbs from CHO cells 60, 61
BioLector	m2p‐labs GmbH	48	0.8–2.4	pH DOT	MTP	Glutathion‐S‐Transferase from E. coli 81
						Organic acids from C. glutamicum 70, 71
						Recombinant parathyroid hormone (rPTH) from H. polymorpha 72
BioLevitator	Hamilton Bonaduz AG	4	50	None	Tube	HeLa cells 174
bioREACTOR 48	2mag AG	48	8–15	pH DOT	STR	C. antarctica lipase B from Komagataella pastoris 64
Bioscreen C Pro	Oy Growth Curves Ab Ltd	200	0.4	None	MTP	L. monocytogenes cell screening 175
						S. cerevisiae cell screening 176
Cellstation	Fluorometrix corp.	12	up to 35	pH DOT	STR	SP2/0 myeloma/mouse hybridoma cell cultures 177
DASbox	Eppendorf AG	24	60–250	Sampler	STR	Cardiac differentiation of human pluripotent stem cells 178
HexaScreen	Telstar Life Science Solutions	6	10–15	Sampler	STR	Suspended and adherent animal cell types 179
Micro‐24	Pall corp.	24	3–7	pH DOT	MTP	CHO cells 180
micro‐Flask	Applikon Biotechnology	up to 96	1	None	MTP	P. putida cultivations 181
micro‐Matrix	Applikon Biotechnology	24	1–5	pH DOT	MTP	CD8 T‐cell line 182
Multifors	INFORS HT	6	100–1000	Sampler	STR	α‐Ketoglutarate from Y. lipolytica 183
						Hydrogen peroxide adapted bifidobacteria 184
SimCell	Seahorse Bioscence Inc.	6 per array	0.7	pH DOT	Microfluidic	mAb from CHO cells 185
Xplorer	HEL	up to 8	1000–4000	Sampler	STR	E. coli fermentation 186

Small‐scale stirred vessels 53, 54, 55, 56 and bubble columns 57, 58 bypass such losses in process information, whereas such systems rather represent a medium‐throughput approach.

The Advanced Microscale Bioreactor (ambr) by Sartorius Stedim Biotech is an ideal HT tool for upstream processing (USP) development as it combines high process information and control as well as low material consumption and culture volume. The ambr system consists of 24 or 48 parallel bioreactors of 10–15 mL, integrated into a fully automated robotic workstation, which enables feeding strategies and sampling. Each cultivation block can be individually tempered, aerated, and mixed while monitoring the oxygen uptake and the culture pH via optical sensors 59. The mixing and aeration is realized by a small‐scale hollow impeller within each bioreactor. The ambr system was used in diverse studies such as in upstream optimizations for mAbs from CHO cells 60, 61 as well as for experiments involving hematopoietic stem cell cultures 59.

A system comparable to the ambr technology allowing for cultivations in the 8–15 mL scale in 48‐well format was introduced by Riedlberger and Weuster‐Botz 62, 63 and commercialized by 2mag as the bioreactor 48 system. Using this system, high‐cell density yeast cultivations for an optimized production of Candida antarctica lipase B was realized 64.

An alternative system, which is frequently used for experiments involving microbial cells, is the BioLector system (m2p‐Labs). The upstream optimization is performed in continuously shaken 48‐well plates, which were geometrically optimized for culture mixing and oxygen supply into the medium 65. The BioLector cultivation plates are all equipped with an optical bottom, enabling quasi‐continuous cell density and product measurements without additional sampling procedures 66. Analogously to the ambr system, such cultivation plates are available with optical sensors for the determination of culture pH and oxygen transfer rates 67. In the RoboLector setup (a combination of the BioLector and a fully automated liquid handling station) as used by Huber et al. 68, different feeding and induction strategies can be realized. Similar experiments were performed by Funke et al. 69, using a microfluidic BioLector derivative. This HT cultivation device can be found widely in the literature, e.g. in product formation studies for Corynebacterium glutamicum 70, 71 and Hansenula polymorpha 72. Scalability of the results in miniature scale can be ensured by the sulfite oxidation method guaranteeing a constant oxygen transfer rate between two scales 73.

3.2. HT cell lysis systems

For bacterial hosts such as Escherichia coli, the commonly intracellular products need to be released from the cytoplasm. As HT cultivation strategies can result in a bottleneck at the product extraction stage, HT‐compatible cell lysis systems need to be considered.

Mechanical cell breakage systems such as bead mills were scaled down to the milliliter scale using beating beads (e.g. NucleoSpin bead system by Macherey–Nagel) agitated by simple mixing. Such a procedure was used for the characterization of bacteria from the human oral and gastrointestinal tract 74, 75.

Also, physical cell lysis systems such as sonication devices were adapted to HT compatibility by using multi‐tipped horns for cell breakage in MTP format. A 24‐well HT sonication device was used by Hohnadel et al. 76 for 15 distinct cell types including gram‐negative and ‐positive bacteria as well as fungi and yeasts. An eight‐well sonifyer for the release of virus‐like particles from E. coli was used by Ladd et al. 77. Another physical HT cell breakage system based on electric fields in parallel microfluidic channels was developed by Poudineh et al. 78. The system introduces strong electric fields into the fluid by multiple sharp‐tipped three‐dimensional electrodes and was successfully applied for extracting intracellular molecules from E. coli. Physical cell lysis in the full 96‐well format can also be realized by straightforward temperature treatment, osmotic shocks, and freeze drying 79.

The full capacities of HTPD in microtiter format can be achieved in terms of chemical (e.g. detergents, solvents, alkaline pH, etc.) and biological (phages, enzymes, etc.) cell lysis. The harvested cells can be re‐suspended in the respective chemicals analogously to the large scale 79. The chemical and biological treatment strategies can be easily implemented into fully automated robotic work stations as shown by Listwan et al. for proteins expressed in E. coli 80. As different strategies for soluble and insoluble protein expression exist, even conditions of IB formation can be identified, e.g. by using trichloro acetic acid–acetone total protein extraction protocols 81.

3.3. HT IB refolding

IBs, being denatured highly pure product aggregates formed in E. coli cells, are a special case in biopharmaceutical process development. Designing IB refolding strategies can be highly feasible especially for rather small proteins such as insulin. The entire process of IB unfolding, protein refolding, and isolation of the correctly folded species is not straightforward, making HTPD the method of choice for designing such processes. An IB refolding process for a fully automated robotic workstation for lysozyme and different proteins from E. coli were introduced by Berg et al. 82, Vincentelli et al. 83, and Dechavanne et al. 84. The protein unfolding and refolding process included investigation of different pH ranges, salts, as well as additives like DTT, glutathione, folding enhancers (chaperones), glycerol, PEG, and co‐factors. In many fundamental studies on protein unfolding and refolding, certain types of additives and chemical buffers were classified as ideal. As a consequence, different commercially available HTPD refolding kits are available such as the QuickFold system (Athena Enzyme Systems) 85, the Pro‐Matrix system (Thermo Scientific) 86, the iFOLD kit (Merck Biosciences) 87, and the FoldIt Screen kit (Hampton Research) 88.

3.4. HT purification systems

Both for intracellular and extracellular products as well as for refolded protein species in a mixture of correctly and falsely folded proteins, a purification strategy is mandatory for meeting the requirements stated by the authorities. The first overview of different HT downstream screening systems and their specifications is shown in Table 2. As many unit operations in the biopharmaceutical industry require defined buffer conditions, different miniaturized devices are available. For buffer exchange and diafiltration, PD desalting columns (GE Healthcare Life Sciences) 89, 90, Slide‐A‐Lyzer Dialysis cassettes (Thermo Scientific) 77, 91, and the Pierce 96‐well Microdialysis Plate (Thermo Scientific) 92 are widely applied in academic research. Vivaspin centrifugal concentrators (Sartorius AG) even enable ultrafiltration and sample concentrations in sub‐milliliter format with different molecular‐weight cutoff membranes 93, 94.

Table 2.

Overview of different commercial miniaturized systems in downstream process development and their applications

Type	System	Distributor	Description	Applications
Clarification	MultiScreen assay system	Merck Millipore	Different filter materials and pore sizes	Solubility assay for pharmaceutical drug candidates 96
	Chromafil Multi 96 filter plates	Macherey–Nagel	Different filter materials and pore sizes	Solubilization and purification of membrane proteins 97
Buffer exchange	PD Desalting Column	GE Healthcare Life Sciences	Bed volume of down to 0.5 mL	Buffer exchange for Mouse Aminoacylase‐3 90
			Down to 130 μL sample volume	Buffer exchange for SF9 insect cell lysate 89
	Slide‐A‐Lyzer	Thermo Fisher Scientific, Inc.	Different Molecular Weight Cutoffs (MWCOs) diafiltration	Diafiltration for virus‐like particle assembly 77 Diafiltration for encapsulation of plasmid DNA 91
			0.1–30 mL sample volume
	Pierce 96‐well Microdialysis Plate	Thermo Fisher Scientific, Inc.	10 kDa MWCO 12 Cartridges of eight dialysis devices 10–100 μL sample volume	Dialysis for cell‐free protein synthesis 92
	Vivaspin Centrifugal Concentrator	Sartorius AG	Different MWCO Ultra‐ and diafiltration 0.5–20 mL sample volume	Ultrafiltration of lysozyme from chicken egg white 94
				Conditioning of human 4‐Phosphopantetheinyl‐transferase 93
IB Refolding	Pro‐Matrix	Thermo Fisher Scientific, Inc.	Refolding buffers customized on protein	Immunotoxins from E. coli 86
	Quickfold	AthenaES	15 different refolding buffers	Cutinase‐like proteins from M. tuberculosis 85
	iFOLD	Merck Biosciences	92 different refolding buffers	Human dihydrolipoamide dehydrogenase from E. coli 87
	FoldIt Screen	Hampton Research	16 different refolding buffers	Bone morphogenetic protein‐2 from E. coli 88
Aggregation	Sirocco Kit	Waters Corp.	Precipitation of 96 samples	Biomarker discovery in biological specimen 187
	RockImager	Formulatrix, Inc.	Crystallization chamber	Crystallization of the Cmi immunity protein from E. coli 188
				Investigating phase behavior of model proteins 99
	Wizard Crystal Screen Kit	Emerald BioSystems	Crystallants, buffers, salts, pH variety	Crystallization of monoclonal IgG4 100
Batch chromatography	PreDictor Plates	GE Healthcare Life Sciences	96‐well pre‐packed filter plates	Binding studies on polyclonal IgG and amyloglucosidase 107
				Binding studies on mAbs 108
	PhyNexus Tips	PhyNexus Inc.	Filter pipet tips pre‐packed with adsorbent	Purification of Fab‐fragments 112
				Crude mixture clearance step for glycoprotein mapping 113
	MediaScout ResiQuot	ATOLL GmbH	Generation of 8 or 20 μL adsorbent plaques	Multi‐component isotherms of modeling quality 110
Membrane chromatography	AcroPrep Advance 96‐well filter plate	Pall Corp.	7 mL membrane volume per well	Purification of granulocyte colony‐stimulating factor 111
	FiltrEx 96‐well filter plates	Corning	Glass fiber membrane for DNA isolation	Separation of hemicellulases in the termite P. militaris 189
Column chromatography	MediaScout MiniChrom	ATOLL GmbH	0.2–10 mL CV	Mixed‐mode chromatography for β‐lactoglobulin from whey 114
				CEX purification of mAbs 115
	HiTrap Columns	GE Healthcare Life Sciences	1 or 5 mL CV	AEX purification of recombinant HIV‐1 Capsid 116
	MediaScout RoboColumn	ATOLL GmbH	0.2 or 0.6 mL CV	BSA and lipolase 117
				HIC of antibody fragments 39
				Miniaturized mAb platform process 118
	AcroSep Columns	Pall Corp.	1 mL CV syringe pumping possible (Luer Lock)	Isolation of an mAb from hybridoma cells 121
	MediaScout CentriColumn	ATOLL GmbH	0.05–0.6 mL CV (centrifugal flow)	New in ATOLL GmbH product line 120
	MediaScout PipetColumn	ATOLL GmbH	0.05–0.6 mL CV (flow by manual pipetting)	New in ATOLL GmbH product line 120

CV = column volume.

3.4.1. HT solubility screenings

Selective protein precipitation is a very straightforward and scalable downstream unit operation for enhancing product purity. As a prerequisite for selectively removing certain species, the solubility lines of products or impurities need to be determined. Different HT screening methods for investigating colloidal stability and protein phase behavior can be found in the literature. Wiendahl et al. 95 developed an MTP‐based method for protein precipitation based on fast liquid evaporation in micro‐scale format. In case studies for lysozyme and insulin variants, systems of different protein and precipitant concentrations were pipetted on a fully automated liquid handling station. After different time frames of liquid evaporation, the turbidity as well as the filling volume were determined until crossing of the solubility line. The presented technology enables the determination of different points within the phase diagram starting from one single starting condition. Different 96‐well filter plates such as the MultiScreen Assay System (Merck Millipore) 96 and the Chromafil Multi 96 Filter Plates (Macherey–Nagel) 97 were used for solubility investigations for pharmaceutical drug candidates and membrane proteins.

An apparatus originally used for screening of crystallization conditions is also used for investigating long‐term protein phase behavior including effects within the metastable and labile regions of the phase diagram. This technology was used for identifying different aggregate states such as precipitate, crystals, skin formation, gelation, and liquid–liquid phase separation for an mAb 98 and different model proteins 99. This methodology can thus not only be used for simple precipitation and solubility screens but can also be seen as a linking technology between DSP development and protein formulation studies. Analogously to the IB refolding kits, also pre‐formulated crystallization kits such as the Wizard crystal screen kit (Emerald BioSystems) used in a case study for IgG4 proteins 100 are available.

3.4.2. HT aqueous two‐phase systems

Another unit operation used in early stages of product purification is aqueous two‐phase partitioning. Two bio‐compatible and non‐miscible aqueous phases composed of salts (phosphate, sulfate), sugars (dextran), or biopolymers (PEG) are mixed with a complex cell lysate. The different biomolecules are then either dragged to the top, bottom, or inter‐phase, which results in a purification effect depending on the distribution behavior. HT technologies can be used for characterization of the aqueous two‐phase systems for the determination of bimodal curves (region of phase separation) and conodes (lines of equal top and bottom phase composition) for different salts and polymers as introduced by Bensch et al. 101. When the systems are characterized, different partitioning studies for biomolecules can follow as shown for model proteins like lysozyme and pharmaceutically relevant molecules such as mAbs 102 and virus‐like particles from human B19 parvo‐virus 103.

3.4.3. HT chromatography

Chromatography is one of the most commonly used unit operations within biopharmaceutical DSP during all stages of process development. Therefore, a large number of screening platforms were developed covering batch resin screenings and dynamic binding systems like miniaturized chromatography columns. Screening for binding and elution conditions as well as comparing different adsorbent media is mostly performed in 96‐well filter plates. Batch binding and elution experiments for kinetic and isotherm studies were applied for mAbs and Fc fusion proteins 104, α‐amylase and scFv‐β‐Lactamase fusion protein 105, and the angiotensin‐II generating enzyme 106 in different IEX and HIC setups in standard filter plates. Pre‐packed PreDictor filter plates (GE Healthcare Life Sciences) are limited to certain adsorbent types and a standard adsorbent volume and were used for bind‐elute screenings for mAbs and amyloglucosidase 107, 108. By contrast, the MediaScout ResiQuot system (Atoll) allows for a flexible generation of equally sized adsorbent plaques for self‐packed adsorbent plates 109. The plaques are generated by applying vacuum to adsorbent slurry, which is sucked into cavities of defined volume. The data generated using this device are of modeling quality as was shown for multi‐component isotherms of lysozyme and cytochrome c 110. Muthukumar and Rathore 111 presented an MTP‐based membrane chromatography approach using AcroPrep Advance 96‐Well filter plates of 7 μL membrane volume (Pall Corp.) for the Granulocyte Colony‐Stimulating Factor, which is an alternative approach to chromatography based on adsorptive beads.

A different system for batch‐binding experiments was commercialized by PhyNexus. Analogous experiments as in filter plates can be realized by simple aspirating and dispensing steps using adsorbent‐packed pipetting tips. Binding, washing, and elution buffers can easily be tested with standard pipetting equipment. The PhyNexus tips are commonly used for fast isolation steps such as affinity chromatography of Fab fragments 112. Prater et al. applied the system as a crude mixture clearance step for glycoprotein mapping 113.

Whereas batch‐binding experiments can be a good start for investigating biomolecule–adsorbent interactions, the industrially relevant dynamic binding setups in chromatography columns are much more complex. Industrial setups are performed far from the thermodynamic‐binding equilibrium with chromatography column contact times being in a range of several minutes only. For generating dynamic binding data in small‐scale and mimicking large‐scale industrially relevant setups, different small‐scale columns were introduced. Miniaturized columns of milliliter scale for LC systems are, e.g. MediaScout MiniChrom columns (Atoll) as used for purification of mAbs and bovine β‐lactoglobulin 114, 115 and HiTrap Columns (GE Healthcare Life Sciences) as applied for recombinant HIV‐1 capsid purification 116.

A further reduction in column size and volume in combination with the possibility to implement column chromatography into the robotic work flow was realized with the MediaScout RoboColumn technology (Atoll) 117. The 200–600 μL columns are not connected to tubings as in traditional LC systems and are placed in a fixing device on the robotic work station. The loading of the columns is performed by application of the liquid handling pipettor into the conically shaped column inlet. Application of different buffers and salt steps requires the aspiration of new liquids from a reservoir, which makes this chromatography technology rather semi‐continuous. The sample collection is realized by catching the droplets leaving the columns in MTPs resulting in a chromatogram resolution dependent on droplet and fraction size. Nowadays, RoboColumns are widely applied in academic research as well as in industry as effective scale down models. Their validity was proven in multi‐stage mAb processes 118, 119, BSA, and lipolase 117, as well as antibody fragment purification 39. Atoll expanded the product pipeline of RoboColumns on miniaturized columns of 0.05–0.6 mL bed volume, which can be operated without the need of a robotic system. One of these systems allows for liquid flow by manual pipetting (PipetColumns), the other by centrifugal force (CentriColumns) 120. AcroSep Columns (Pall Corp.) are another straight‐to‐use column chromatography system without the need of robotic or LC equipment used for the purification of mAbs from hybridoma clones 121. The columns are equipped with a Luer lock system enabling a direct connection to syringes for liquid flow.

4. Modeling approaches

Gaining a more profound understanding of biopharmaceutical USP and DSP is a key demand of the Quality by Design guideline. Where heuristic knowledge and trial and error HT technologies describe a good foundation for process understanding, modeling evolves as a central tool in process development 122. Modeling and simulations can drastically reduce the number of experiments while enhancing or capturing the experimental content 122. Models are generally divided into two categories: empirical and mechanistic models. Empirical models are based on a priori known output data (e.g. certain results from HT experiments) within a defined design space. The created model then allows for predictions for unknown experiments within the calibrated or extrapolated region without mechanistic or biochemical foundation of the model parameters. Per contrast, mechanistic models are based on physicochemical properties and rely on calibration experiments for the determination of model and property parameters. Osberghaus et al. 123 compared both types of models in an IEX chromatography case study for the separation of lysozyme, ribonuclease a, and cytochrome c revealing the limitations of simple empirical models.

4.1. Response surface modeling

RSM is one example of empirical modeling that fits a regression function for experimental results collected under defined conditions without the knowledge of the physicochemical background of the estimated parameters 124. Generally, linear or quadratic response surface models are widely used for IEX chromatography, whereas models of higher order are known to be more and more error‐prone 123. RSM is common practice in Design of Experiments (DoE) programs such as Modde (Umetrics), Design‐Expert (Stat‐Ease Inc.), Fusion Pro (S‐Matrix Corporation), or the JMP software (SAS Institute), which is widely applied in pharmaceutical industry for investigating the influence and interference of different parameters such as pH, ionic strength, etc. on different quality attributes 125. DoE‐RSM is employed at all stages of biopharmaceutical production: In USP applications, laccase production in Panus tigrinus cultures was modeled with respect to the impact of parameters like glucose and nitrogen concentration and three different inducers 126. Islam et al. 127 created an RSM model on soluble recombinant protein production in MTP format using E. coli under different media, growth, induction, and aeration conditions yielding in a three‐fold increase in firefly luciferase concentration. In DSP applications, DoE‐RSM was used for designing an alternative mAb mixed‐mode chromatography purification step including variation in chromatography media, pH, and conductivity 128. A detailed discussion on bioprocess optimization strategies based on empirical models was reviewed by Mandenius and Brundin 124.

4.2. Multivariate calibration modeling

RSM is regarded as black box modeling, which is why the trends discovered can be error‐prone and misleading. Also, data generated in a biopharmaceutical process or derived from computational simulations are often so complex and multitudinous that RSM fails to find truly predictive models. Multivariate calibration modeling (MCM) describes an efficient way to analyze multi‐dimensional datasets using projection methods such as principal component analysis and projection to latent structures 129. Complex datasets can then be condensed to the factors of highest information content. Based on training sets, the multivariate calibration models can be calibrated and validated using cross‐validation or internal test sets 130.

The application of MCM in biopharmaceutical manufacturing is diverse. Currently, it is used for the collection of real‐time process information in terms of multivariate data analysis of spectra. Vaidyanathan et al. 131 used near‐infrared spectral data for investigating complex feedstocks from an antibiotic production process. Deconvolution of protein spectra in the mid‐UV region was shown to be highly useful in determining protein concentrations from multi‐component mixtures 132, 133, 134. Brestrich et al. 135 expanded the application of such a methodology to industrial process steps including in‐line quantification of higher and lower molecular weight species during mAb polishing and of serum proteins during the purification of Cohn supernatant. Principal component analysis of mid‐UV protein spectra was also employed for investigating conformational stability of lysozyme under high‐salt HIC conditions 39. Diverse applications of MCM used in biomedical analysis were reviewed by Escandar et al. 136. Xu and Glatz 137 used projection to latent structures for correlating the pI, molecular weight, and the aqueous two‐phase partitioning coefficients of different proteins to their retention behavior on IEX chromatography media.

Another hot topic in biopharmaceutical research is the combination of MCM and molecular dynamics simulations called quantitative structure‐property (QSPR) or structure‐activity relationship (QSAR). Starting from in silico three‐dimensional molecule structures under defined chemical conditions, different property and structure descriptors such as hydrophobicity, charge distribution, and size are extracted into a data bank 9. Experiments performed under defined conditions then serve as a training set for correlating the most influential molecular descriptors to certain molecule properties or activities. The selection and collinear linkage of the descriptors is based on MCM as discussed above. The validated QSAR or QSPR models can then be applied for extrapolation and predicting the behavior of new molecules or process conditions. Yang et al. 138 used QSPR for characterizing multi‐modal IEX ligands with respect to protein binding under high‐salt conditions, as demonstrated for the model proteins RNase A, cytochrome c, and lysozyme. Schaller et al. 139 used the three‐dimensional structure of proteins for the prediction of soluble peptide surfactant DAMP4 variant expression in E. coli. In a study by Asadollahi‐Baboli and Mani‐Varnosfaderani 140, different thiazolidine‐4‐carboxylic acid derivates were investigated in terms of influenza virus neuramidase inhibitor activity and correlated to the respective molecule structure in a QSAR approach. Dismer et al. 141 used molecular descriptors such as the relative hydrophobicity of PEG molecules to predict phase separations and even partitioning coefficients for lysozyme as a model protein. MCM can thus be a link between empiric and purely mechanistic modeling and the validity of the models relies on the choice of calibration experiments and the correct data handling.

4.3. Mechanistic modeling

Whereas empiric models as in RSM simply describe the ultimate outcome of a process under defined conditions (black box model), mechanistic models are based on a fundamental understanding of the underlying mechanisms resulting in the observed results. Mechanistic models thus need to include physicochemical properties of different interacting species to describe the final outcome. Such models are mostly of the convection diffusion reaction type describing hydrodynamics (e.g. equations of continuity, mass balances), mass transfer (e.g. film diffusion), and interaction (e.g. protein–protein, protein–ligand). In a simplified case of batch thermodynamic equilibrium, a mechanistic model can be of the reaction type only. Andrews and Roberts 142 described protein phase transitions such as non‐native protein aggregation in the Lumry‐Eyring Nucleation‐Polymerisation model. The thermodynamic model includes reversible conformational changes of the monomers, reversible non‐native oligomer formation, as well as irreversible conformational changes during nucleation and aggregate growth. Also, thermodynamic models for protein–ligand interactions as used in chromatography are widely applied in pharmaceutical industry and research. Limousin et al. 143 reviewed different adsorption isotherms, which are commonly used depending on the complexity of the protein‐binding process. The list of binding isotherms includes models such as linear, Freundlich, Temkin, and Langmuir in different combinations as well as in single‐component and multi‐component format. Foo and Hameed 144 extended the list of binding isotherms including models such as the BET isotherm, which accounts for multilayer binding after coverage of the adsorbent surface. The common single‐component Langmuir isotherm depends on two estimation parameters only, being the equilibrium‐binding coefficient K _eq and the maximum binding capacity q _max. When assuming multi‐layer adsorption as for the BET isotherm, an additional solute solubility factor c _sat is included. The complexity of sorption models can be increased steadily, whereas the number of parameters to be estimated rises accordingly and might result in over‐fitting. Different attributes included in multiple isotherms such as linear and non‐linear adsorption, salt ion dependence, and induced salt gradients upon binding are discussed by Iyer et al. 145. The semi‐mechanistic steric mass‐action isotherm introduced by Brooks and Cramer 146 as an adsorption model for IEX chromatography includes factors for varying salt concentrations, counter‐ion displacement upon protein binding, as well as characteristic protein properties such as the characteristic charge and steric shielding. The steric mass‐action model was applied for the AEX binding of rotavirus‐like particles 147, and the separation of mAb monomers and dimers 145. Sellberg et al. 148 used a reaction model that combined both, a chromatography model for mAb binding behavior and a coupled model for product dimer formation.

When moving from pure reaction models to kinetic models including hydrodynamics and mass transfer phenomena (convection diffusion reaction type), different modes of operation can be chosen while keeping the model as simple as possible and as detailed as needed 149. In experimental setups of high mass transfer resistance, the general rate model (GRM) is a suitable option. It covers fluid convection, axial dispersion of the fluid, as well as internal and external mass transport phenomena 150. The GRM was employed for modeling the IEX separation of BSA and IgG 151 as well as for extracting B19 parvovirus‐like particles from an Sf9 insect cell broth using a hydrogel‐crafted AEX membrane 152. The transport‐dispersive model is a simplification of the GRM as film and pore diffusion is combined in a lumped effective mass transfer coefficient. It is commonly used for modeling standard column chromatography procedures. Such a setup was employed for the AEX separation of a liver phase‐2 enzyme produced in E. coli 153 and modeling of an antibody platform process 148. Another lumped‐rate model including convection, dispersion, and adsorption kinetics is the reactive dispersive model 154. A further reduction in complexity yields in a model such as the equilibrium dispersive model (EDM), which only accounts for convection and axial dispersion, while neglecting internal and external mass transfer. Close et al. 155 used an EDM for optimizing operating conditions of an HIC setup to purify a therapeutic protein from a crude feed stream. The reduction in model complexity resulted in drastically reduced calculation times while assuring the identification of critical quality attributes. A simplification of the EDM results in the ideal model that excludes axial dispersion effects and only accounts for convection 154.

Different in silico optimization studies can be found in the literature including separation of small molecules 156, model proteins 123, mAbs 157, 158, as well as industrial process applications 159, 160. Besides, custom‐designed tools for modeling also commercially available solutions exist, making in silico tools accessible for a broader public. Such programs include Aspen Chromatography (AspenTech) 161, CADET (Forschungszentrum Jülich) 162, Chromulator (Ohio University) 163, ChromWorks (Ypso‐Facto) 164, and ChromX (Karlsruhe Institute of Technology) 165.

5. Combinatorial approaches

Approaches based on heuristics, HT experimentation, and modeling alone constitute a good basis for fast and effective process development, but their full potential can only be realized in a combination of these above‐mentioned procedures in so‐called combinatorial approaches. Such a combination of approaches can be in terms of multi‐stage or superstructure optimizations of consecutive unit operations, or in terms of complex feedstock characterization for tailored process development. The latter can either be realized experimentally (deterministic approach) or in silico (hybrid approach) as illustrated in Fig. 2. A large number of lysates generated in miniaturized cultivations under varying growth conditions (blue box) can be analyzed in terms of critical impurities (CIs) either experimentally (red box) or using computational methods (yellow box) for fast and dedicated downstream process development. CIs are generally regarded as molecules of similar or identical physicochemical properties as the product. Feasible lysates from product cultivation are analyzed with respect to their impurity profiles resulting in an integrated evaluation of USP and DSP performance (Fig. 2, center). Feasible process conditions can then either be used as an optimized process or experimentally refined by further exploration of the design space.

Figure 2.

Schematic illustration of an integrated biopharmaceutical process development for up‐ and downstream applications based on combinatorial approaches. A large set of feasible complex feedstocks from micro‐scale cultivations are characterized in terms of critical impurities. The characterization can be performed experimentally (deterministic approach—red box) or in silico (hybrid approach—yellow box).

5.1. Multi‐stage/superstructure process development

Multi‐stage or superstructure process optimization does not aim at optimizing single unit operations but at finding an ideal combination of consecutive steps. As the number of variables increases exponentially with the number of possible process stages, such approaches can hardly be realized experimentally. Computational approaches have the highest potential for performing a maximum number of in silico experiments. Samsatli and Shah 166 presented an integrated design for biochemical processes by a two‐stage optimization. The first stage represents the dynamic phase for the optimization of volume handling, operational constraints, processing times, and the number of maximum unit operations. The second stage then accounts for time schedules and total consecutive process design 167. Asenjo et al. 168 proposed a strategy for designing a batch protein plant in a modular structure using process performance models and keeping plant structure as well as process variables flexible. Vásquez‐Alvarez and Pinto 169 presented mixed integer linear programming models for designing consecutive chromatography process development. Each specific chromatography model includes physicochemical properties of the molecule pool and possible purification routes to guarantee economic processes. In a case study by Huuk et al. 170, a purification procedure for two consecutive IEX steps was optimized in an integrated manner for a three‐component mixture of proteins. The study was performed by in silico modeling of both setups in all possible combinations while the output pool of the first column served as the input signal of the following step. Zhou and Titchener‐Hooker 171, 172 presented the technology of integrated bioprocess design based on operation windows and multi‐objective Pareto optimization. The sweet spots of different unit operations can be analyzed graphically to make decisions on feasible combinations as shown for an integrated optimization of cell homogenization and debris removal of an intracellular Saccharomyces cerevisiae enzyme process.

5.2. Deterministic approach

The deterministic approach is a combination of the experimental and the heuristic approach. Well‐chosen screening experiments serve for determination of the key parameters of certain processes and complex lysate characterization. For IEX chromatography, pH gradient chromatography was proposed as a tool for identifying CIs with similar elution conditions as the product. Ahamed et al. 173 investigated the elution behavior of 17 model proteins including related species such as isoforms in terms of elution pH and nominal pI from the literature. Kröner et al. 32 performed a multi‐dimensional analysis for the isolation of Nucleolin from Sf9 feedstocks starting a pre‐fractionation using pH gradients. In a study by Baumann et al. 7, pH gradients served as a systematic screening tool for IEX process development and lysate characterization including salt‐intolerant proteins as demonstrated for human α‐Galactosidase A from a complex Pichia pastoris supernatant. For HIC, such a characterization based on pH gradients is hard to accomplish. However, the central parameter as an analogue to the elution pH in IEX was identified as the protein solubility 39. Conditions of decreased solubility result in an increase in HIC binding. As a consequence, CIs in HIC are supposedly molecules of similar phase behavior as the product.

5.3. Hybrid approach

The hybrid approach can be seen as a link between experimental process development and modeling. The determination of key parameters and CIs for complex lysate characterization is performed in silico. As mechanistic models allow for modeling of the product as well as lumped impurity peaks in different chromatography setups, even peaks overlapping with the product can be identified. Cell lysates from different cultivation conditions can thus be analyzed in silico using computational sampling of all possible fractionation setups followed by multi‐objective characterization. Baumann et al. 47 used such a setup for an integrated USP and DSP optimization for a phase II liver enzyme from E. coli. It was shown that the overall optimal system must not be the upstream condition of highest initial product titer.

Hanke and Ottens 9 reviewed different setups that can be used in hybrid approaches in terms of HT data generation and in silico predictions such as molecular properties (QSPR, QSAR, molecular dynamics) and model parameters such as isotherm data. Such properties include molecular mass, conformation, diffusivity, net charge, hydrophobicity, solubility, interaction coefficients, and many others.

6. Concluding remarks

With the biopharmaceutical industry going besides large‐quantity productions such as antibody platform processes to personalized or precision medicine, new fast and effective processes need to be developed within a short time frame. The Quality by Design guideline stated by the authorities demands a deeper process understanding, which is why traditional heuristics alone are no longer accepted. The implementation and commercialization of HT and mechanistic modeling tools at almost all stages of biopharmaceutical productions have opened up an entirely new playground for process development.

This overview summarizes the pillars of process development of heuristic, experimental, or model‐based nature. It is shown how simple heuristics can help find suitable initial datasets for experimental exploration of the design space and how experimental data can inversely result in new platform processes. Also, with mechanistic models becoming the gold standard in the biopharmaceutical industry, the experimental approach will always be needed for model calibration and validation purposes. Finally, the synergetic power of combinatorial approaches is shown in deterministic and hybrid approaches. Experimental or in silico lysate characterization enables up‐ and downstream optimization in an entirely integrated fashion.

With such new tools, it will now be possible to react on changes in the biopharmaceutical workflow. This will be especially important for continuous processing setups, which are subject to changes in the feed stream over the production time.

The authors have declared no conflict of interest.